The Impact of Education on Family Formation: Quasi

 The Impact of Education on Family Formation:
Quasi-Experimental Evidence from the UK*
Michael Geruso
University of Texas at Austin & NBER
[email protected]
Heather Royer
UC Santa Barbara & NBER
[email protected]
This version: July 21, 2014
ABSTRACT
In this paper we examine the fertility and mating market effects of education. We exploit a quasiexperiment generated by a change in UK compulsory schooling laws. This change, introduced in
1972, forced all students to stay in full-time education until at least age 16. The reform was recent
enough that access to legal abortion and modern contraception was quite similar to today, granting
insight into the fertility effects of education in a modern context. This reform was binding for many
girls, inducing around one quarter of the female population of England and Wales in the relevant
cohorts to attend an additional year of school. For identification, we leverage the fact that
compulsory school requirement was discontinuous with respect to cohort of birth using regression
discontinuity methods. We show that the affected girls had significantly lower fertility in their teen
years. Instrumental variables estimates imply a 30% reduction in births at ages 16 and 17 caused by
the additional year of schooling. The decline was not accompanied by any increase in abortions. We
also find that the reform had negligible impacts on completed fertility. Our findings suggest that
education-based policies might reduce teen pregnancies without impacting completed fertility rates.
On the mating market front, the reform induced both men and women to marry more educated
mates and induced women to marry younger mates. The mating effects suggest that educational
reforms can have multiplicative effects on household income by changing mate quality.
We thank Julie Jefferies and Louise O’Leary for the considerable time and effort that they and their team have devoted
to this project. It would not have been possible without the live births dataset that they supplied. We also thank Mary
Grinsted for generously supplying the abortion data used in this paper and Kevin Lynch for his help with the
*
1
1 Introduction
Beginning with Becker (1960), economic analysis of fertility and family formation has been of
perpetual interest to researchers. Of particular interest has been the relationship between female
education, labor, and fertility decisions. The early literature (e.g. Becker 1960; Mincer 1963; Becker
1965; Willis 1973) was motivated in part by the extremely robust negative correlation between
individual-level completed fertility and a woman’s educational attainment, with attention to the
opportunity cost of childbearing. Most recently, policy interest has focused on two topics related to
fertility, education, and labor: First, on fertility trends towards below-replacement levels in highincome countries (e.g. Kalwij 2010); and second, on the levers for moving teen fertility, including
and especially education.1 Teen fertility is an outcome widely considered bad for both teen mothers
(Geronimus and Korenman, 1992; Goodman, Kaplan, and Walker, 2004; Chevalier and Viitanen,
2003; Ashcraft and Lang, 2006) and their children (Royer, 2004; Levine and Painter, 2004). The
fertility effects are only one of a host of important family formation effects of education, which
include effects in the mating market (Behrman, Rosenzweig, and Taubman, 1994).
Despite wide interest in the relationship between education and family formation, evidence of causal
effects remains under-developed. While partial correlations between education and various fertility
and marital market outcomes raise the possibility that education affects family formation, they do
not necessarily imply a causal connection. The omitted variables problems that are pervasive
throughout empirical studies are likely prominent here. For example, a female from a disadvantaged
family is at higher risk of a teenage pregnancy (Geronimus and Korenman, 1992), as well as at a
higher risk of early exit from school.
In light of the strong correlation between education and other characteristics influencing fertility, it
is not surprising that the empirical evidence on the education-fertility relationship is inconclusive.
Black, Devereux, and Salvanes (2008) uncover large teen fertility effects exploiting compulsory
schooling reforms in the US and Norway. But in the long-run, the effects on completed fertility in
Norway are much smaller (Monstad, Propper, and Salvanes 2008). In contrast, McCrary and Royer
(2011) find no effects when they utilize variation in educational attainment due to school starting
rules in the US. Currie and Moretti (2003) estimate significant fertility responses to college openings
but have crude measures of fertility.2
This literature suffers from several weaknesses. Nearly all of this literature relies on difference-indifference analyses, which can often be sensitive to the specification of coincident time trends.
Other studies of compulsory schooling reforms often involve the estimation of the effects of several
See for instance, the website for the National Campaign to Prevent Teen Pregnancy
(http://www.thenationalcampaign.org).
2Within the economics literature, empirical work on the fertility effects of education is more complete for developing
countries, where recent large national pushes to increase education can be used to identify the fertility effects of
education. Examples include Breierova and Duflo (2004), Osili and Long (2008), and Duflo, Dupas, and Kremer (2011).
These still tend to focus on teen outcomes, in part because the affected cohorts are too young to have completed their
fertility.
1
2
educational reforms (like those in Scandinavian countries) as countries often change compulsory
schooling requirements along with implementing other educational reforms. In such situations, it is
difficult to evaluate the impact of extending the length of schooling in isolation.
In this paper, we estimate the causal effect of education on fertility (timing and quantity) and the
mating market. Our analysis can shed valuable new light on this empirical question for several
reasons. First, we exploit sharp and plausibly exogenous variation in education. Specifically, we use a
1972 UK education reform that raised the minimum school-leaving age in England and Wales from
15 to 16. The change affected all women born in September 1957 or later. As such, we can rely on
regression discontinuity (RD) comparisons of women born very close in time (e.g., September 1957
versus August 1957) who might be otherwise comparable aside from their education. In other quasiexperiments, RD estimators are not often feasible, either because of the resolution of the data or
because the implementation details don’t permit it (e.g., in the United States, Lleras-Muney 2005).
Second, the reform impacted a large fraction of the population: roughly one quarter of the women
born in September 1957 received an additional year of schooling as a result of the reform. The
power of this quasi-experiment is observed in Figure 1, which plots month-of-birth averages of the
probability of dropping out by certain ages along with regression fits to those averages. The fraction
of individuals dropping out by age 15 falls by roughly 30 percentage points as a consequence of the
compulsory schooling age increase. The strength of this fall allows us to estimate the effects of this
reform very precisely. It also means that our local average treatment effects are estimated for a large
and policy-relevant group, somewhat in contrast to settings such as the US, where compulsory
schooling laws have been binding for only a few percent of young people (Goldin and Katz, 2003).
Third, the reform was recent and thus the institutional features, such as access to legal abortion and
modern contraception, were quite similar to today, allowing us to examine these effects in a modern
institutional context. Yet enough time has elapsed since the reform for us to explore the completed
fertility effects.
Fourth, in comparison to the prior literature, we can better address the effects of education on
conceptions, as opposed to solely on births, by examining impacts on abortions. We combine
information on educational attainment from the UK Labor Force Survey with administrative records
on the universe of live births and abortions in England and Wales. Ours is the first quasiexperimental study of the fertility effects of schooling that includes abortion outcomes.
Finally, by exploiting a unique feature of a complimentary survey, we can directly identify a subset of
compliers affected by the reform. Their self-reported attitudes and partnership histories during their
teen years suggest that the mechanisms driving the teen fertility effects may arise in the mating
market, rather than as changes in behavior with the same types of partnerships. We investigate this
in an RD framework, using Census data on partnerships.
Our analysis reveals four main findings. First, an additional year of schooling had a significant
impact on teen fertility, reducing it by around 10% at ages 16 and 17. Second, the additional
schooling had at most a negligible effect on teen abortions. This implies that the teen fertility effects
3
reflect a reduction in conceptions. Third, we cannot reject that an additional year of schooling had
no impact on post-teen fertility and no impact on completed family size. 3 Fourth, the extra
schooling led affected individuals to choose more educated partners, with women marrying men
closer to them in age.
We discuss several explanations for why this educational reform could have reduced teen fertility.
Crucially, we show that the timing of these teen fertility reductions is consistent with education
reducing fertility by reducing the probability that girls conceive while they are in school. We also
discuss why this additional education appears not to have changed total fertility. From a theoretical
perspective, the most obvious explanation for the zero effect on total (completed) fertility is that the
income and substitution effects of the additional earnings enjoyed by more-educated women roughly
cancelled out. While Willis (1973) suggests that the substitution effect is more likely to dominate,
that analysis ignores the effects of women’s education on husband’s income, a mechanism discussed
by Becker (1981) and documented by Lefgren and McIntyre (2006). With regard to teen fertility, our
findings suggest that it was the period of schooling itself, rather than longer-lived human capital
effects of education that impacted childbearing behavior.
With the important caveat that we study only one margin of education, the effects of which may be
different from those at other margins (e.g., years in college), we conclude that there is no evidence
that increased education necessarily leads to lower fertility. Therefore, increasing educational
attainment at the population level in the OECD need not imply reductions in future levels of
completed fertility. This is a particularly important implication for Western Europe where low
fertility is a major concern in light of pension systems and public health systems and where low
fertility motivates significant policy intervention (see Kalwij 2010 for an overview). Nevertheless,
our study does suggest that schooling can be an effective tool in the fight against teen fertility in
developed countries.4 This is important both because teen fertility reduction is often a difficult
policy goal to achieve and because, as discussed below, policy-makers in the UK, US, and other
countries are currently considering measures that would compel girls to stay in school for longer.
The UK has recently extended the compulsory schooling age to 17 and will increase it to 18 next
year. Moreover, extensions of the compulsory schooling age have added benefit of augmenting
family resources directly through the direct effect of one’s own education but also that of one’s
mate.
3
While the teen fertility reductions are large as a percentage of teen births, teen births are such a small part of overall
births that complete pass through of the teen effects into completed fertility would fall well within the tight confidence
intervals of the completed fertility effects.
4Teen fertility reduction is often cited as an important motivation for increasing educational attainment among women in
developing countries.
4
2 Compulsory Schooling Laws and Fertility Trends
2.1 Compulsory Schooling Laws
Compulsory schooling laws in the UK are set at the national level.5 They include a maximum age by
which children must start school and a minimum age at which children can leave school. Officially,
children must be age 5 by September 1st of their year of entry; this has been stable since the late
1800’s.6 In contrast, the minimum leaving age has changed four times in the postwar period. These
changes, also known as ROSLAs (Raising of the School Leaving Age), were implemented in 1947
and 1972. The ROSLA was further increased in 2013 and is slated to change in 2015.7
The 1947 ROSLA increased the minimum leaving age from 14 to 15. Several papers have used this
reform to study the effects of education on various outcomes (e.g., Harmon and Walker, 1995;
Oreopoulos, 2006; Devereux and Hart, 2010; and Clark and Royer, 2013). The more recent 1972
ROSLA increased the minimum leaving age from 15 to 16. This second ROSLA has received less
attention in the literature; Harmon and Walker (1995) and Grenet (2013), Clark and Royer (2013),
Machin et al. (2011), and Oreopoulos (2007) use it to study the impact of education on wages,
health, crime, and happiness respectively. We focus exclusively on the 1972 ROSLA because fertility
data are not available for the women affected by the 1947 ROSLA.
The 1972 ROSLA applied to students in England, Wales and Scotland. We focus on the effects for
women born in England and Wales because we do not have fertility data for Scotland. During
World War II, the government legislated both the 1947 ROSLA but also recommended a schoolleaving age change to 16 when feasible. 8 Later, the Crowther Report and the Newson Report
recommended ROSLA to 16 in 1959 and 1963, respectively. Finally in 1964 action was taken to
implement the change; the slated date was 1970-1971. By 1968, they postponed the ROSLA for two
years. In March 1972, it was officially announced that the school-leaving age would increase by one
year. The new school leaving age became effective on 1 September 1972.
Students who turned 15 before September 1972 (born August 1957 or earlier) were subject to the
old law; students who turned 15 after this date (born September 1957 or later) were subject to the
new law. In the Appendix we discuss in more detail some important aspects of the 1972 ROSLA.
A detail particularly relevant for our study is the treatment of pregnant females within the schooling
system. Both before and after the ROSLA, girls who became pregnant during the period of
compulsory schooling were required to continue with education through their pregnancy. They
Much of this discussion is adapted from Clark and Royer (2013).
Local Education Authorities have allowed children to enter school before 5, but the frequency of this practice has
diminished over time (Del Bono and Galindo-Rueda 2007).
7 The Education and Skills Act 2008 stipulates that individuals remain in schooling or training until the age of 17 starting
in 2013 and until the age of 18 beginning in 2015. See website: http://www.politics.co.uk/reference/education-leavingage.
8 These details come from Woodin, McCulloch, and Cowan (2013).
5
6
5
attended school until the later months of their pregnancy, after which they received an in-home,
state-provided tutor. Thus a teen pregnancy did not result in a waiver from the compulsory
schooling requirement.
2.2 The Impacts of the 1972 ROSLAs
There is a plethora of possible mechanisms behind effects of the 1972 ROSLA on family formation.
We highlight a few here. First, an increase in a woman’s education could impact tastes for
childbearing directly. Second, education could alter the labor market opportunities—through
occupation or through wages—which could change the demand for children and the desired timing
of childbearing. Third, education could influence the rate of time preference and thereby change the
timing of childbearing. Fourth, the extra schooling could impact health and thereby affect fertility
through both intensive and extensive margin. Fifth, it could influence the choice of mate, either
through the mechanisms described above or directly through tastes.
Prior research, albeit limited, gives us some insights as to which of these channels may be operating
as a consequence of the ROSLA.9 The instrumental variable estimate of the return to an extra year
of schooling for the ROSLA-impacted cohorts is 6-7 percent for both men and women (Grenet
2013). These cohorts also are less likely to receive welfare and are happier (Oreopoulos 2007).
Health impacts of the additional education were minimal (Clark and Royer 2013).
3 Data
Our study uses several datasets. We describe each of these in turn. The common variables linking
our datasets are the month and year of birth of women whose education, fertility, and partnership
outcomes we examine. Our analysis is at the level of month-of-birth cohorts, as we describe later. A
summary of the data sources and the relevant variables we extract from each is provided in
Appendix Table 1.
UK Labor Force Survey. Information on educational attainment comes from the 1975-2002 UK
Labor Force Survey (LFS). The LFS samples a large number of households to provide information
on labor force participation, employment, and training. Around 60,000 households are interviewed
in each wave. Between 1975 and 1983 the LFS was conducted biannually. Between 1983 and 1992
the LFS was an annual survey of a similar number of households. Since 1992, the survey has been a
rotating quarterly panel, with the same households interviewed for five consecutive quarters. In the
analysis, we include multiple observations per person when available, and cluster all results at the
cohort level. Respondents’ year and month of birth are available in public release LFS datasets up to
2002.
Since its existence, LFS respondents have been asked about the age at which they first left full-time
continuous education. Respondents were also asked about the highest academic qualifications they
held, though the exact form of these questions differs across LFS survey years. Our pooled sample
9
Most of the literature using the UK compulsory schooling laws focuses on the 1947 ROSLA not the 1972 ROSLA. 6
of 1975-2002 LFS gives us a large sample of over 300,000 respondents born between 1954 and
1962.
Live Birth Records. Live birth data were supplied by the UK Office for National Statistics (ONS).
The records cover every birth between 1970 and 2008 in England and Wales to mothers born in
England, Wales, or Scotland – roughly 600,000 to 900,000 births per year.10 Since teen fertility is a
relatively rare outcome, it is crucial to have a census of births. The births data are birth counts by
mother’s month and year of birth and child’s month and year of birth but do not include parity
information. Given that we rely on across-cohort variation in education, except under special
circumstances, it is not problematic that we do not have individual-level birth data. Specifically, our
regression framework, described later, does not require the inclusion of covariates aside from
controls for cohort trends. Any pre-determined covariates, in principle, should be similar for females
born immediately prior to September 1957 and females born immediately following September 1957
because sorting around this threshold would be impossible to do since the exact timing of the
ROSLA change was announced in 1972.
The structure of our data implies that we can recover cohort-specific fertility rates but not
individual-level fertility rates. For instance, our data cannot distinguish between a scenario in which
one woman bears her first child at age 20 and another bears her first child at age 24 and a scenario in
which one woman bears her first child at age 20 and then her second at age 24. In practice, since we
find that the reform only impacts fertility among girls aged 16 and 17, who are very likely to have
been first-time mothers, the missing parity information is of little consequence. Moreover, assuming
that education affects fertility behavior mainly in one direction (i.e., either increasing it or decreasing
it), the inability to distinguish parity will not cause us to mischaracterize the results.11
For each cohort of women, we construct age-specific and total fertility rates according to standard
practice (Preston et al. 2001). The numerator in the age-specific rate for birth cohort c is the number
of births to women of cohort cover a particular age range (e.g., [16, 17)). The denominator is the
number of women alive in that age range. Unfortunately, it is not possible to know exactly how
many women are alive at each point in time. Thus, we generate the denominator by multiplying the
size of the female cohort at birth (obtained from the Registrar’s Annual Reports) with a succession
of age- and calendar year-specific survivor rates for the UK (obtained from the Human Mortality
Database). We validate these denominator estimates by comparing our predictions for each cohort
size in 1991 and 2001 with cohort size as measured by the 1991 and 2001 Censuses. Our estimates
and these numbers line up very closely. This correlation ranges between 0.93 and 0.95. The total
fertility rate (TFR) aggregates the age specific rates. Cohort TFR reflects the average number of
The birth records from England and Wales do not distinguish between mothers born in England and Wales and
mothers born in Scotland. This poses minimal problems for our study, since only a small fraction of the women living in
England and Wales during this time were emigrants from Scotland (<1%).
11 For example, counter to this assumption, suppose that the reform influences one woman to increase her fertility
preferences from 2 to 3 children and influences another to reduce her preferences from 1 child to being childless. In this
case, the total number of children born is unchanged but this overall effect masks important heterogeneity.
10
7
children that women in the cohort bear over their lifetime, conditional on surviving through their
fecund years.
Abortion Records. Detailed abortion records come from the UK Office of National Statistics. Like
the live birth records, these data are counts by women’s month and year of birth. Unlike the live
birth records, these are limited to a narrower range of cohorts born around the reform threshold.
They also cover only abortions performed in the teen years (the only years for which we find any
fertility effect). These data allow us to construct teen (16-19) abortion rates for cohorts of women
born within a 36-month window of the ROSLA threshold.
UK Census Extracts. Finally, we commissioned several extracts of the 2001 Census that are useful
for understanding the mating market effects of the extra education. These extracts allow us to track
the age difference between partners as a function of each partner’s month of birth and the
educational credentials of each spouse in relation to their own month of birth.
National Child Development Study. Finally, we use data from the National Child Development
Study (NCDS), a panel that attempted to track 100% of babies born in Great Britain in a particular
week of March 1958. This cohort was one of the first to be subject to the 1972 ROSLA.
Information collected at ages 7, 11, 16, 23, 33, and 41 offer insight into cohort members’ school
activities, partnerships, and labor force participation during the extra year of compulsory schooling
and beyond. Interestingly and arguably unique to this study is our ability to identify compliers and
always takers in these data. When participants were 16, the study asked both the cohort members
and their parents when they would have dropped out of school had the 1972 ROSLA not been in
place.
4 Empirical Strategy: Regression Discontinuity Framework
Our empirical strategy leverages the difference in the school exit age resulting from the 1972
ROSLA. Specifically, we use a regression discontinuity framework to compare individuals born
immediately on either side of the September 1st 1957 cutoff, who we argue are similar except for
their relevant ROSLA. Our analysis proceeds in two steps: first, documenting that the 1972 ROSLA
indeed did change educational attainment as intended, and second, examining the subsequent effects
the ROSLA had on fertility and family formation.
Educational Impacts of the 1972 ROSLA
To estimate the first-stage educational impacts of the 1972 ROSLA, we follow the fuzzy regression
discontinuity setup (Lee & Lemieux, 2010; Imbens & Lemieux, 2008). That is, we estimate local
linear regressions equations of the following form using data near the September 1957 threshold:
Ec = γ0 + γ1Dc + γ2Rc + γ3DcRc + uc
(1)
8
where Ec is an average educational attainment measurement for month-of-birth cohort c (e.g., the
fraction leaving school at age 15 and the average age at school exit), Dc is an indicator for whether
the cohort is affected by the 1972 ROSLA (i.e., equals 1 for cohorts born September 1957 and later
and 0 otherwise), Rc is the running or forcing variable (i.e., calendar month of birth), and uc is the
error term.12 While the Labor Force Survey data on education would allow us to estimate this
regression using the individual-level micro-data, all of our outcomes (births, abortions, education
and age of mate) are measured at the month by year-of-birth level, so we aggregate our data to the
month by year-of-birth level. We weight these regressions by cohort size calculated as described in
the data section. The inclusion of Rc captures cohort trends, and the inclusion of DcRc allows the
cohort trends to differ on either side of the September 1957 although in principle, we have no a
priori reason why they should, other than that these cohorts experienced their later teen years during
a time of rapidly changing fertility trends throughout the OECD.
Our primary interest is in estimates of γ1, the effect of the ROSLA on educational attainment. This
parameter can be thought of as the “jump” in the dependent variable at the program threshold. This
is the sense in which our research design exploits the comparison of cohorts “just unaffected” and
“just affected” by the reform.
Fertility and Mating Market Impacts of the 1972 ROSLA
Similarly to evaluate the effects of the ROSLA on our main outcomes of interest—fertility and
assortative mating outcomes—we estimate regressions of a form very similar to that of equation (1)
except that our dependent variable has changed. That is,
Yc = β0 + β1Dc + β 2Rc + β 3DcRc + vc
(2)
where Yc is an outcome of interest, β1 in the regression tables can be directly interpreted as estimates
of the ROSLA impact. Note that in none of the employed datasets do we have information on both
educational attainment and a fertility or mating market outcome. Thus, to derive the return to an
extra year of schooling on any of our outcomes (e.g., fertility at a certain age, educational attainment
of mate), we can estimate β1/γ1, the standard Wald IV estimate using a two-sample instrumental
variables strategy (Angrist and Krueger, 1992).
Choice of a Bandwidth
To estimate these equations, we use data close to the September 1957 threshold. We must decide on
how close to the threshold we should be (i.e., bandwidth). To date, there is no universally accepted
bandwidth selection procedure. We follow the cross-validation procedure of Ludwig and Miller
(2007).13 The method minimizes the square of the prediction errors from the local linear model.
12 Note it is not immediately apparent that educational attainment can be immediately derived from age at school exit.
However, for most, educational attainment can be derived from age at school exit with little error because retention rates
are low (<5%). See http://eacea.ec.europa.eu/education/eurydice/documents/thematic_reports/126EN_HI.pdf
13 Other bandwidth selector procedures are Imbens and Kalyanaraman (2012) and Calonico et al. (2014).
9
Because we estimate a variety of effects (births at 16, 17, 18, 19, etc.), for each collection of
outcomes (e.g., education or births), we choose one bandwidth for each similar set of outcomes. For
example, when looking at teenage fertility, we find the strongest visual evidence of effects at age 17,
and choose the bandwidth that performs best in the cross-validation statistic for the age 17
outcome, and then use it for all of the other teenage fertility outcomes. For robustness, we estimate
the regression discontinuity models with both narrower and wider bandwidths.
One other common approach in empirical applications of the regression discontinuity design is the
global polynomial approach. That is, instead of focusing narrowly around the threshold, this
approach uses more flexible functions of the running variable (i.e., higher order terms of R). We
show below that cohort patterns in our data are highly irregular away from the threshold, and global
polynomials do a poor job of fitting the data near the threshold.
Identifying Assumptions
A key advantage of the regression discontinuity approach is the minimal set of assumptions needed
for identification. We must assume that individuals born immediately prior to and immediately
following September 1957 are otherwise identical except for their educational attainment. Ex ante,
this assumption seems plausible as the policy change was not pre-announced. Moreover, it is
testable, at least in terms of observable characteristics. That is, for each observable characteristic, we
can test whether the characteristic has a discontinuity at the September 1957 threshold. Clark and
Royer (2013) examine number of births and stillbirth rates and find no evidence of discontinuities in
these outcomes.14
An identifying assumption under which our IV estimates are valid measures of the effect of a year of
additional schooling at age 15 is that the 1972 ROSLA does not affect fertility through channels
other than its effect on the probability of experiencing a year of full-time education at age 15. This
exclusion restriction would be violated, for instance, if the law was accompanied with a change in
the accommodations for pregnant students.
5 Education Results
In this section we report our main results. We begin by reporting estimates of the impact of the
ROSLA on educational attainment. We then report estimates of the impact of the ROSLA on
fertility and mating market behavior.
5.1 Education
To establish that the ROSLA impacted educational attainment, Figure 1 shows the impact of the
1972 ROSLA on age at school exit for women. Specifically, each point in this figure represents the
There is a little set of pre-determined characteristics to examine since education could affect many of them. Moreover,
we view behaviors as adults and very few datasets that sample adults have parental information.
14
10
average outcome for a cohort of women who share the same month and year of birth. The three
sets of points trace the fraction of women who left school by age 15, by age 16, and by age 17. The
lines are local linear regression fits to those underlying points.
The compulsory schooling law was binding for a significant share of individuals. Almost one third of
women born before September 1957 left school at age 15. The reform had an immediate impact on
age at school exit as there is a clear drop in the fraction leaving before age 16 (i.e., before grade 10 in
a US context) starting with the September 1957 cohort. This gap is on the order of 25 percentage
points. Effects for men (not pictured) are similar. However, there are minimal impacts observed for
leaving school by age 16 and 17.
While Figure 1 suggests that most students complied with the new law, there is a non-negligible
share of individuals who report not complying (<8% for most cohorts). The most apparent noncompliance is a seasonal pattern of among cohorts born in June, July, and especially August. This is
because school was not in session when these women turned 16. Hence, they report leaving school
at 15, even though they are likely to have completed grade 10, in compliance with the law, and left
school in practice at age 15.7 or later. For non-summer borns, it appears that much of the reported
non-compliance is measurement error. Because the LFS is a panel, respondents are asked to recall
their school leaving age in multiple waves, and over 10% of respondents who report leaving school
at age 15 also report leaving school at 16 in a different survey wave. Further, in the NCDS, which
interviewed members of one of the first post-reform cohorts when the respondents were 16 years
old—therefore eliciting responses much closer in time to the school-leaving event—less than 1%
reported leaving school before age 16.
In Table 1, we quantify the patterns observed in Figure 1. In Panel A, we report discontinuity
estimates of γ1 for different educational outcomes from estimation of equation (1). The estimation
sample includes 48 months of birth cohort data on either side of the ROSLA threshold (i.e., 96
months in total). The reported mean of the dependent variable is for the August 1957 birth cohort.
The ROSLA led to a 27 percentage point reduction in the fraction of individuals leaving schooling
by age 15. This is a considerable change given that the fraction of the last pre-ROSLA cohort
leaving at age 15 was 0.44.
As evident in Figure 1, the ROSLA did not induce large fractions of women to stay in school
beyond the extra year. Specifically, the effects on leaving at age 16 and 17 are -0.02 and -0.01.
Overall the effect on age at left full-time education is 0.31 of a year, which is highly statistically
significant with a t-stat of over 5. This discontinuity is noticeable in Figure 2, which plots the
average age at school exit by birth cohort. These first-stage estimates are essentially identical to those
of Clark and Royer (2013).15
15 In fact, all of the point estimates are the same except for that for left school by age 16, which is -0.037. Clark and
Royer (2013) use different data (the General Household Survey and the Health Survey of England) and a different
bandwidth.
11
While these regressions verify the existence of a first-stage relationship, several other points from
Panel A are worthy of discussion. First, the size of the impacted population is considerable. Over a
quarter of the post-reform cohort had its school leaving age impacted by the 1972 ROSLA. As such,
the estimated local average treatment effects (LATE) may be closer to the average treatment effect
(ATE) in this context (Oreopoulos 2006). For comparison, compulsory schooling changes only
impacted 5 percent of population in the US (Lleras-Muney, 2005) and only around 12 percent in
Norway (Black et al. 2008). Second, the set of compliers for whom we will be identifying the effect
of a year of education is a policy-relevant subpopulation. Third, we do not have to worry about weak
instruments given the strength of the first stage. Such worries have hindered analysis of compulsory
schooling in other countries (see Black et al (2008) for a discussion of this).16 Also, somewhat unique
to this context, the strength and robustness of the first stage allows us to use regression
discontinuity methods rather than difference-in-difference methods; the regression discontinuity
design may be preferable given the relatively weak assumptions needed for identification.
As noted above, one of the regularities in Figure 1 is the seasonal pattern in the probability of
leaving school, particularly by age 15. To address this, in panel B, we re-estimate regression equation
(1) and include a full set of calendar month-of-birth dummies excluding the September month-ofbirth dummy.17 Given the consistency of this pattern across cohorts, these calendar month-of-birth
dummies may adequately control for these regularities. Also, for further robustness, instead of
calendar month-of-birth dummies, we exclude the summer borns (i.e., those born in June, July, and
August) in estimation of equation (1) in Panel C. The results are similar to those from Panel A
assuring us the strength of this research design. We have also tested for sensitivity of the education
effects to varying the bandwidth (see Appendix Table 2). These results are consistent with those in
Table 1.
6. Fertility-Related Results
6.1 Teenage Fertility
We start with examining the effects of the ROSLA on teen fertility.18 Figure 3 presents plots of agespecific fertility rates in a format analogous to that in Figure 2. These fertility rates are calculated
according to age-last-birthday. Thus the age 16 rate captures births in the exact age range [16,17),
and the age 17 rate captures births in the exact age range [17,18). The magnitude and significance of
these ROSLA effects is explored in Table 2, which presents reduced-form estimates on teen fertility
from ages 15 to 19.19,20
16 Black et al (2008) show that the addition of state-specific trends along with the appropriate calculation of standard
errors to difference-in-difference regression models estimating the effect of compulsory schooling on education using
US data leads to weak first-stage estimates.
17 We have also estimated models where we allow the effects of calendar month of birth to differ on either side of the
September 1957 threshold. The results are similar. 18 General fertility patterns in the UK and Europe are discussed in the Appendix and help to give some context to the
underlying cohort trends we observe here.
19 Our data do not allow us to examine fertility at ages below 15. The administrative birth records begin in 1970, which
means that the first cohort for whom we could calculate an age 14 birth rate is January 1956. This leaves us with too
12
Looking at panel A of Table 2, the ROSLA led to significant effects on age 16 and age 17 fertility.
Since these effects are somewhat sensitive to specification, it helps to view these estimates in the
context of Figure 3. Both the effects on age 16 and 17 fertility are visually apparent in Figure 3.
These effects are also economically meaningful. A reduction of -0.0018 in age 16 fertility
corresponds to a 10 percent decline. After scaling by the first stage effect on education of 0.31, this
represents a 32 percent decline. For age 17 fertility, the reduced-form effect is an 8 percent fall
implying a 26.3 percent IV effect.
The effect on age 16 fertility varies a bit across the alternative specifications in Panels B & C
although the local linear fits from Panel A apparent in Figure 3 seem to be adequately capturing the
patterns in the data. The IV effect for age 16 is still sizable according to these alternative
specifications - a 16 percent decline. On the other hand, the age 17 fertility effect is more consistent
across the other specifications. These estimates are also robust to different bandwidths (see
Appendix Table 2).
Since the ROSLA kept girls in school between ages 15 and 16, and since ROSLA fertility effects are
concentrated between ages 16 and 17, it is important to pinpoint the exact timing of the ROSLA
fertility effects and, ultimately, the exact timing of the conception effects that precede them. In a
later section we address this question by estimating teen fertility effects at narrower age ranges - for
example between ages 16 and 16 and one quarter. These analyses also help to determine whether
there were statistically significant ROSLA effects on fertility prior to age 17. In relation to this
question, the evidence presented in the first three columns of Table 2 is somewhat inconclusive,
despite the visual evidence in Figure 3.
Beyond age 17, we observe no ROSLA impacts on teenage fertility; the reduced-form effect sizes are
2.6 percent and 0.3 percent for ages 18 and 19 fertility.21 These small impacts are evident both in
Figure 3 and Table 2. There does, however, appear to be a significant change in the slope of the age19 fertility trend near this threshold. Upon closer inspection the timing for this change would have
been 1976-1977. This time effect, which occurs in the age 19 plot among cohorts born in late 1957,
is also apparent in the age 18 plot among cohorts born in late 1958, in the age 17 plot among
cohorts born in late 1959, and in the age 16 plot among cohorts born in late 1960. Since the graphs
plotted in Appendix Figure 1 reveal that both France and Germany shared the same trend break, it
seems safe to conclude that this has nothing to do with the ROSLA but more to do with general
trends in fertility at that time.
small a pre-reform sample to do meaningful estimation. However, little is lost by this limitation. Fertility at age 14 is a
rare event, even in our dataset of the universe of births. Thus any age 14 estimates would have been associated with large
standard errors and driven by just a handful of births.
20 The Table 2 teen fertility estimates are based on a narrower data window than the Table 1 education estimates. These
different windows are those suggested by a cross-validation procedure. Intuitively, the fertility window is narrower
because, around the ROSLA threshold, fertility trends are more rapidly changing than education trends. In suggesting a
narrower window, the cross-validation procedure implicitly takes account of this feature of the data.
21
The finding of no ROSLA effects on fertility at ages 18 and 19 extends to ROSLA effects on fertility rates in one-year
age intervals up to 45. 13
In terms of understanding the nature of these teenage fertility effects, we can further divide the
outcome into in-marriage and out-of-marriage births. Appendix Figure 2 shows the regression
discontinuity plots for in-marriage and out-of-marriage birth rates; Table 3 presents the analogous
regression discontinuity estimates with standard errors. Appendix Table 2 examines the robustness
of these estimates to varying bandwidths.
The figure clearly shows that the fertility impacts we see for ages 16 and 17 come from a decline in
in-marriage fertility. In this population and time period, essentially all marriages among 16-17 yearold girls were preceded by a conception. Because all births in-marriage were indicative of a shotgun
marriage in this sense, one implication of the observed drop in in-marriage births is a corresponding
decline in shotgun marriages. There is no fall in birth rates for out-of-marriage fertility. Not
surprisingly, the magnitudes of the regression discontinuity are similar to those in Table 2 for fertility
overall. It is difficult to surmise without additional information whether for this age group, the
welfare of the parents and child are better if the birth occurs in or out of marriage. It is also not
possible to determine whether, within the context of the observed overall decline in births, there was
substitution towards out-of-marriage status.
6.2 Abortions
A missing element in past education and fertility studies is how abortion may mediate the impact of
education on fertility. Usually abortion data are difficult to come by. For example, in the United
States abortions are not universally reported across states, but in the UK, due to the existence of the
National Health Service, which provides universal health insurance, abortion statistics are considered
more reliable and comprehensive. We also think that the context of the 1972 ROSLA is unique
because unlike other compulsory schooling studies, the quasi-experiment we exploit occurred in the
era of modern fertility when abortion and contraceptives were available. Abortions are an important
part of fertility; the ratio of abortions to births in our sample varies from 1.7 for 15 year olds to 0.30
for 19 year olds. The teenage abortion rates experienced by the post-ROSLA cohorts are similar to
that experienced by cohorts today. Additionally, information on abortions allows us to get closer to
understanding how education affects conception and sexual behavior than looking at the effects of
education on fertility alone.
To see the importance of observing abortions, consider that the negligible effects of education on
fertility for the older teenagers may reflect similar rates of contraception of pre- and post-ROSLA
cohorts or higher rates of conception among pre-ROSLA cohorts with higher rates of abortion for
these cohorts. Similarly, the education-related reductions in fertility among 16 and 17 year olds
would be consistent with either with a difference in conception rates among pre- and post-reform
cohorts or with comparable conception rates but a higher rate of abortions among the post-reform
cohorts due to the higher opportunity cost of having a birth for these cohorts.
Now we turn to the abortion data to test for these intervening effects. In Table 4, we present the
reduced-form regression discontinuity estimates for abortions at age 15 through 19; Figure 4 shows
the analogous figures. Across all of the ages and all of the specifications, there is no strong evidence
14
of an impact of the ROSLA on abortion rates. These conclusions are robust to the alternative
bandwidths (see Appendix Tables 2). Just examining the point estimates, most of the reduced-form
estimates range 1 to 3 percent. The exception is the age 15 abortion rate, which implies a 5 to 9
reduction in abortion rates, but because the incidence rate is so low, the estimates are imprecise.
6.3 Conceptions
Rather than examining abortions and births separately as a means of estimating the effects on
conceptions, we can combine these data together to estimate a conception rate. To do this, we
adjust the timing of abortions to be consistent with that of births. For example, an abortion in
September 1974 would have likely resulted in a birth in the first half of 1975 if the abortion had not
occurred. To do so, we move the timing of abortions forward by 7 months (most abortions occur
before 12 weeks of gestation). Then, we simply add abortions and births together to form our
“conception” rate. Of course, this measure misses miscarriages.
Appendix Table 3 presents the ROSLA effects on our measure of conceptions. The increase in
school exit age reduces the conception rate at ages 16 and 17 (12 percent and 9 percent reducedform effects). These effects are statistically significant at the 1 percent level in Panel A. These
magnitudes align with those for birth rates. Effects for ages 18 and 19 are much smaller and
statistically insignificant.
6.4 Cumulative Fertility
Since the earliest of the ROSLA-impacted cohorts have completed their fertility by the early 2000’s,
we can also examine the impacts on completed family size. To begin, Figure 5 plots the regression
discontinuity figure for cumulative fertility from ages 16 to 45 (inclusive). Cohort-specific cumulative
fertility is often referred to as the cohort total fertility rate (TFR). 22 This cumulative rate is
constructed as the sum of each cohort’s age-specific rates. Because of the nature of our births data,
we do not observe completed fertility at the person level. However, cohort TFR would exactly equal
the average completed family size found in individual data provided that all women in the cohort
survived through their childbearing years. Since the probability of survival from 16 to 45 among
these cohorts was above 98 percent, we view our cumulative fertility measure as a good proxy for
average completed family size.
Figure 5 suggests that there was no ROSLA effect on completed fertility. The scale of the vertical
axis measures average births per cohort member, and the apparent zero effect of the reform is
contained within a very narrow band of births per woman. Table 5 presents regression discontinuity
effect sizes and standard errors for a expanded set of cumulative fertility outcomes. Fertility between
16 and 25, 16 and 30, 16 and 35, and 16 and 40, and 16 and 45 are considered as outcomes. The
reduced form effects implied by the graph and in Table 5 range from -0.0055 to 0.0073 children.
Even after this reduced form effect is scaled by the ROSLA effect on education, this range is only Cohort TFR, which is derived from age-specific rates for a fixed cohort, differs from period TFR, which is derived
from age-specific rates for a fixed point in time.
22
15
0.018 to 0.023 children. To put this effect size in context, note that for cohorts of women born near
the reform threshold, an additional year of education at age 16 is associated with a decline in total
childbearing of 0.248 children in the cross section (see Geruso 2010). Our IV estimates are an order
of magnitude smaller. The finding of no completed fertility effect is robust to alternative regression
specifications; see Appendix Table 2 for results under alternative bandwidths.
In even the simplest economic models of fertility, fertility and educational attainment are governed
by the same joint decision process. These depend, among other things, on preferences, family
background and peer influences, all forces that researchers cannot easily measure. These results
suggest that the strong correlation found between education and completed family size reflects the
impacts of this basic heterogeneity.
At first glance, it may seem that our findings of large negative teen fertility effects and small
completed fertility effects imply positive “catch up” effects beyond the teenage years. This is not
necessarily true. Instead, because teen fertility is such a small component of completed fertility, the
confidence intervals on our fertility effects include both this catch-up story and the polar opposite
story under which the teen fertility effects lead to a permanent (albeit tiny) reduction in completed
fertility.
These possibilities can be seen in Figure 6, which aggregates information from 30 distinct regression
discontinuity estimates from local linear regressions that include month of birth dummies. Each
point represents an estimate of the cumulative fertility effect from 16 up to the indicated age. The
left panel displays the regression discontinuity estimates when cumulative fertility is measured in
levels, while the right panel uses cumulative fertility measured in logs. As fertility rates vary
considerably with age, the log graph estimates can be interpreted as percentage change. Since the
first point is degenerate, representing the cumulative effect from 16 to 16, it corresponds to the
ROSLA effect on the age 16 fertility rate (from column (2) and panel B of Table 2). The figure
shows that the cumulative effects begin at age 16 and then die out, as the averted teen births become
swamped by the larger number of births at older ages. This is consistent both with teen effects being
counteracted by post-teen effects and with alternatively, teen effects generating permanent
reductions in completed fertility.
7 Understanding the Mechanisms Behind the Fertility Effects
Classical theories linking fertility and education emphasize the role of opportunity costs (Becker
1960 and Willis 1973). Increases to a woman’s education raise the opportunity cost of her time and
lead to the combination of income and substitution effects influencing fertility in opposite
directions. To explain the strong cross-sectional correlation between education and fertility, Willis
(1973) argues that the substitution effect trumps the income effect. Becker (1965) and Becker and
Lewis (1974) discuss a substitution from quantity towards quality with ambiguous effects on family
size. Note that under any such theories, one would expect the influence of education to be
permanent (i.e., impact fertility throughout the life cycle) and not short-lived in the teen years as we
observe here.
16
The temporary nature of the effect we find is more consistent with an alternative role of schooling
for impacting fertility. In particular, the teen fertility reductions appear to follow the compulsory
schooling period with a lag of about one year, suggesting a connection between the fertility effects
and the new period of schooling itself. It seems as though being enrolled in school changes the
opportunity cost of time, the ability to find a mate, or other factors related to being in school. The
findings of other studies confirm this role for education. For example, in Chile, the lengthening of
the school day is associated with fertility reductions (Berthelon and Kruger 2011).
7.1 Detailed Timing
To explore the idea of this incapacitation effect of schooling, we now examine the timing of fertility
effects relative to the timing of the new compulsory schooling period in more detail. Figure 7 tests
this conjecture that the reduction in pregnancies carried to term is influenced by current school
attendance. Each panel plots ROSLA impacts on fertility rates defined over a succession of threemonth intervals. The leftmost point on each plot corresponds to ages 15-15.25, the second point to
ages 15.25-15.5 and so on. The associated confidence intervals are represented by the dotted lines
around each of these points. The vertical dashed lines depict the new compulsory schooling period
shifted forwards by nine months to indicate where to expect a reduction in births if conceptions
were reduced during the new compulsory enrollment. For most individuals, this additional
compulsory schooling occurred between exact ages 15.5 and 16.5.23 Although the correspondence is
not exact, the figure shows that roughly speaking, the ROSLA effects are largest 9-12 months after
the new schooling period.
To explain the teen fertility effects then, we must explain why girls were less likely to become
pregnant during the new compulsory schooling period. An obvious explanation is that they had
fewer opportunities to become pregnant: less contact and less time with older men that they might
have met in other contexts, such as the workplace. We can observe very little about the partnerships
of teen girls, but aggregate national statistics from England and Wales show that most marriages
among teen girls in this time period (which would be comprised primarily of shotgun weddings for
16 and 17 year olds) are made with men two or more years older. To shed light on the possible
incapacitation story in which schooling kept girls separated from older potential partners, it would
be useful to know more about the types of people that ROSLA-affected girls associated with and
their preferences over older partners. For that, we turn briefly to a complimentary panel survey that
followed one of the first post-reform cohorts.
23 Unlike in the US, students in the UK wishing to leave school at the earliest permissible date could not do so
immediately upon their 16th birthday. In most cases, they had to complete a school term that ended some time after
their 16th birthday.
17
7.2 Direct Evidence Partnerships from Complier Self-Reports
We gain insight into preferences over mates and fertility using the age-16 and age-23 waves of the
NCDS. All NCDS respondents were born in March 1958, making their cohort one of the first to be
affected by ROSLA. The NCDS data is unique in that it contains self-reports of the desired schoolleaving age. At the time they were 16, cohort members were asked: "Do you wish that you could
have left [school] when you were 15?" We use the answer to determine the desired age for leaving
school. In the case of missing or uncertain responses to this question, we use the parents' answer to
the question: "In the study child’s case do they wish that he/she had been able to leave school at
fifteen?" We use this information, combined with other survey responses to provide intuition for the
potential mechanisms behind our results.
Adopting the classic instrumental variables language, we call respondents who wished to leave
school at 15 but stayed on until 16 the compliers, and the respondents who left at 16, but didn’t
wish to leave at 15, the always-takers. Because these are hypothetical and not actual responses, we
assess the possibility validity of these responses by comparing the fraction of NCDS compliers
(0.27), to the fraction of LFS respondents in the pre-reform cohorts who report actually leaving
school at 15 (0.34). As further evidence that self-reports are informative of the counterfactual in
which the school-leaving age had not been raised, we show that regional variations in stated
preferences (post-reform) and realized schooling outcomes of earlier cohorts (pre-reform) are
strongly correlated. Appendix Figure 3 is a scatterplot showing that the fraction of NCDS
respondents wishing to leave school at age 15 in each region highly correlates with the fraction of
LFS respondents actually leaving school at age 15 in that region in the pre-reform period. The
correlation is among female respondents is 0.72, despite that the LFS measures region of current
residence and the NCDS measures region of birth.
Panel A of Table 6 lists some descriptive statistics of the female NCDS cohort members, stratified
by complier status. First, it is interesting to note that at age 16, after the additional year of schooling,
the compliers and always takers differ in their preferences over mates and childbearing, with
compliers preferring earlier marriage and earlier children. Almost 20% of compliers preferred to be
married by 18, compared to just 11% of always-takers. We do not know whether the additional year
of schooling had no effect on the childbearing preferences of the compliers, but we do know from
these statistics that experiencing the additional year of schooling did not cause to their preferences
to converge to those of the always-takers, who also completed their schooling at age 16.
Panel A also shows that a significant fraction of girls in both groups worked during full-time
schooling. Compliers were slightly more likely to do so and worked more hours. Work is a plausibly
important context for defining peers. Because the respondents in the NCDS are all post-reform,
compliers and always takers do not differ in the compulsory schooling age they faced. However, the
high prevalence of work in both groups, even during full-time education, suggests that education
could have been a binding constraint on the labor force activities of compliers. If so, compulsory
schooling could have altered compliers’ exposure to a different set of peers and potential mates.
18
We can look more closely at certain kinds of partnerships in the NCDS to gain insight into mate
characteristics. Panel A shows that while essentially all women lived with their guardians on their 16th
birthday, a substantial minority left home at 16 or 17, many to begin cohabitating with a mate. In
Panel B of Table 6, we explore the characteristics of cohabitating partnerships that began at these
ages. Among both compliers and always-takers, the men these women partnered with were
substantially older. Among 16 year-old women, the 25th percentile of the partnering man’s age was
18 and the median was 20. A full third of the women cohabitating 16 had a birth 16, and over one
half had a birth by 17. Similar patterns are found for women who began cohabitating at 17. These
descriptive statistics show that women who were in serious relationships at 16 and 17 were
uniformly partnering with much older men.
The fact that both compliers and always-takers lived at home until the end of schooling, after which
a substantial minority began cohabitating with older men, suggests that the reform may have been
delayed the initiation of cohabitating partnerships for complier women. Thus one potential reason
that fertility declined at ages 16 and 17 is that the additional year of schooling limited women’s
exposure to older men at age 16. For a more systematic analysis of this hypothesis, we use census
data below to explore the effect of the reform on the probability of marrying a similarly-aged
classmate, which could serve as a substitute for an older man. We also investigate effects on the age
of spouse. Estimates of both effects are consistent this hypothesis, with women marrying men closer
to them in age.
There are several other factors that could explain why girls were less likely to become pregnant
during the new schooling period, although none are entirely convincing because they do not predict
temporary decreases in fertility. First, the extra schooling may have weakened girls’ preferences for
motherhood. It is not hard to imagine why education might have such an effect, but it is hard to
explain why this type of effect would be temporary. That is, why would it not reduce fertility at later
ages? Second, the extra schooling may have provided girls with better information about birth
control. Again though, one would expect girls to retain this informational advantage beyond the new
compulsory schooling period, resulting in reduced fertility at later ages, including at 18 and 19.
8 Mating Market Results
Aside from fertility, another aspect of family formation potentially impacted by education is
assortative mating (Lefgren and McIntyre 2006). We will look at assortative mating over the
dimensions of education and age. In the presence of assortative mating, the ROSLA could have a
multiplier effect – for example, affecting both the education of the woman and the spouse she
chooses.24
8.1 Spouse’s Education
For some UK evidence on this question, Figures 8 & 9 and the corresponding Table 7 provide the
ROSLA impacts on partner’s educational level. Figure 8 shows the probability of the male spouse
24
The ROSLA could also impact a man’s education and that of his spouse.
19
having no qualifications, 1 or more O-level exam or 1 A-level exam, and 2 or more A-level exams or
higher as a function of the wife’s date of birth. These categories are mutually exclusive. O-level and
A-level exams are sets of national exams covering different subjects. O-level exams happen after
grade 9 (the grade in which most will turn 16). A-level exams are administered two years later. Figure
9 is the analogous figure for men. It is important to keep in mind that we observe these individuals
in 2001, many years after the ROSLA. As such, we interpret these effects as the long-run impacts on
mate choice. Since all of these figures are conditional on marriage, in analysis not reported, we also
look at the probability of marriage as a function of date of birth. There are no impacts on marriage
probabilities for either gender.
Both women and men affected by ROSLA were less likely to marry men or women with no
qualifications. The regression discontinuity estimates are similar for men and women – a statisticallysignificant reduction of 1-1.5 percentage points. Figures 8 & 9 visually confirm these estimates.
These estimates are remarkably stable across specifications, including varying bandwidths (Appendix
Table 2). When scaled by the first stage estimates, these effects amount to a 15 to 20 percentage
point reduction for women and men, respectively, in the probability of marrying an individual with
no educational qualifications. Given conventional estimates of the rate of return to schooling, and
assuming that the difference between no qualifications and some qualifications corresponds to a year
of schooling, this spousal education channel could lead to important effects of education on
household income.
It should be noted that the population affected on the mating market dimension is not necessarily
the population affected on the fertility dimension. Many more individuals have their mating market
decisions impacted than their fertility decisions (i.e., 1 percentage point mating market effect versus
0.03 percentage point age-16 fertility effect).
Instead of marrying mates with no credentials, the ROSLA-impacted cohorts married men or
women with some O-level qualifications/1 A-level qualification; we observe small and insignificant
impacts on marrying someone with 2 or more A-level qualifications or higher.
8.2 Spouse’s Age
Schooling could impact the educational attainment of one’s mate in several ways. An extension of
schooling could influence the composition of their age 15 peer group and make one more likely to
marry someone whom they met through school. Additional education could change spousal
preferences or opportunities to meet more-educated potential husbands or wives (e.g., through
changes in occupation or employment status).
For further insight into the mechanisms behind the education effects, we now turn to examining
how the ROSLA impacted spousal age. On average, men in this context marry women who are
roughly 2 years younger. Thus, if the increase in required schooling increases the probability of
marrying a schoolmate, we would expect that men and women would marry individuals closer in
20
age.
In Figure 10, we display the regression discontinuity graphs for the age difference. In the left panel
we show the husband’s age minus wife’s age as a function of the woman’s month of birth. In the
right panel we show the wife’s age minus husband’s age as a function of the man’s month of birth.
These plots are seasonally-adjusted by removing month-of-birth effects.25 For women, there is a
sharp decrease in the average age gap at the threshold consistent with an effect of the ROSLA. That
is, women who had to stay until 16 marry younger men. The corresponding regression discontinuity
estimate is -0.87 of a month (roughly a 3 percent reduced-form effect), which is statistically
significant at the 1 percent level.26 On the other hand, there is not a similar effect for men.
The average age graphs obfuscate interesting underlying distributional effects. In Figure 11, we plot
the regression discontinuity estimates for each integer age gap along with 95 percent confidence
intervals, adjusting for seasonality via the inclusion of month-of-birth controls. The top panel
displays that for the age gap of husband’s age-wife’s age as a function of wife’s date of birth and the
bottom panel displays that for the age gap of wife’s age-husband’s age as a function of husband’s
date of birth. For the top panel, the 0.004 at age gap at 0 implies that the women born immediately
following September 1957 are 0.4 percentage points more likely to marry a man between 0 and 1
year older than them than women born immediately prior to this date.
For women, the impacts imply that the ROSLA increased the probability of marrying a man 1 and
11 years older and 3 years younger and reduces the probability of marrying a man 2, 3, 8, and 9 years
older. One plausible characterization of these patterns is that the extra compulsory schooling raises
woman’s likelihood of marrying a much older man but also shifts some women who have married
men 2 to 3 years older to marrying men 0 to 1 years older.
For men, the shifting of the age gap is much more modest. As a result of the ROSLA, men are less
likely to marry women who are 2 years younger but more likely to marry women who are 3 years
younger or the same age as themselves.
We can test more directly whether the education assortative mating effects are due to the fact that
extending the length of schooling changes the peer group that one is exposed to. We do this in two
ways. First, in Appendix Figure 4, we compute the fraction marrying someone in their academic
cohort and see whether we observe a regression discontinuity in that outcome. An academic cohort
starts with September births and ends with August births of the following year. There is
considerable but systematic seasonality in the probability of marrying a classmate. For women, the
oldest in class are less likely to marry a classmate than the youngest in the class (a 0.11 probability
versus a 0.13 probability). For men, the pattern is the reverse with a slightly larger difference
25 We regress the age gap on a set of month-of-birth dummies, take the residuals, and add back the mean age gap to
obtain these seasonally-adjusted numbers.
26 This estimate comes from a local linear regression with month-of-birth dummies.
21
between the oldest and the youngest (a 0.15 probability versus a 0.11 probability). This seasonality is
unaffected by the ROSLA. Due to this seasonality, the Appendix Figure 4 plots the de-seasonalized
patterns to be able to visually discern the ROSLA effect. Both for men and women, the probability
of marrying a classmate increases because of the ROSLA. Using the specification that controls for
month of birth (i.e., Panel B specification in the regression tables), the regression discontinuity point
estimates are 0.0042 and 0.0032, respectively for women and men – both are statistically significant
but small.
Second, it might be possible that the assortative mating effects arise due not to marrying someone in
one’s own class but someone who was also subjected to the 1972 ROSLA. In Appendix Figure 5, we
calculate the fraction marrying someone born September 1957 or later. This raw figure exhibits little
seasonality so we present the figure unadjusted for seasonality. For both men and women, there is
an apparent discontinuity in the probability of marrying an individual whose compulsory schooling
age was 16. Local linear regression regression discontinuity estimates (not reported) including
month-of-birth dummies are 0.069 for women and 0.054 for men. At least some of the spousal
education effects result from ROSLA-affected women being more likely to marry ROSLA-affected
men and vice versa, irrespective of their preferences for spousal education or opportunities to meet
more-educated potential husbands or wives. Policies that change one gender’s education levels
without changing those of the other would likely have weaker effects. Nevertheless, theory and
evidence suggests education effects on spousal education and it is interesting to see those effects in
this context.
9 Discussion
Given the strong negative correlation between education and fertility often found in observational
data, it is somewhat surprising that we find no evidence of an education effect on completed fertility.
It is unsurprising that OLS estimates are biased towards large fertility effects - OLS estimates are
unlikely to control for all of the factors that might generate a spurious correlation between education
and fertility. This is particularly true since many of these are hard to measure, such as women’s
attitudes towards their role in society. But our results and standard errors for completed fertility
(through age 45) rule out even a small fraction of the OLS effects. On face value, this contradicts
some theories of the education-fertility relationship, which predict that education reduces fertility via
its impact on wage rates and hence on the opportunity cost of raising children.
In relation to theory, note that if children are a normal good, then other things equal, we would
expect more-educated women to demand more of them. Willis (1973) acknowledges this, but argues
that these positive effects are likely smaller than the negative effects operating via the opportunity
cost mechanism. In the context of his model, this conclusion seems reasonable. This model does,
however, impose some strong assumptions. Once these are relaxed, the positive income effects may
be as strong as the negative opportunity cost (substitution) effects.
22
One important assumption is that women must choose between working and raising children. An
alternative formulation might have women facing a labor market penalty for raising children, where
this penalty varies across time, across countries and across occupations. If public policies and work
practices have become more family-friendly over time, this penalty—and thus the strength of the
opportunity cost mechanism—will have weakened.
The problem of teen fertility is one that attracts a lot of attention. Discussions of how to reduce it
have focused on sex education, the availability of contraception and the way the welfare system
treats teenage mothers. In this paper we have shown that it might be reduced by policies that result
in girls spending more time in school. Such educational policies have other very positive benefits
including improving labor market opportunities and thus likely provide a high return on investment
in comparison to other policies aimed at decreasing teenage pregnancy.
The decline in completed fertility is a related problem that policy-makers have grappled with.
Education does not feature in this policy discussion because, if anything, the conventional wisdom is
that more-educated women have fewer children. Our analysis suggests that there is no causal
relationship between education and completed fertility in the modern, developed country context we
examine: Our estimates are statistically indistinguishable from zero and our confidence intervals
restrict the possible effects to a small range.
Aside from fertility decisions, this reform had other effects on family formation through positive
assortative mating. ROSLA-impacted women and men married more educated mates. For women,
the age gap between spouses fell with women marrying younger mates. At least some of these
mating market effects are attributable to being in school as both men and women became more
likely to marry classmates and more generally others who were also impacted by the 1972 ROSLA.
This education assortative result is particularly important as the target population of the ROSLA
includes those who are risk of dropping out of school and thus, those who may have reduced means
as adults. The ROSLA has a multiplier effect through impacting both the husband’s and wife’s
education level, further reducing the likelihood that this at-risk population lives in poverty.
Our findings are important as policy-makers in both the UK and US are consider measures that
would result in girls spending more time in school. The UK Coalition Government increased the
compulsory schooling age to 17 in September 2013 and will further augment it to 18 by 2015. Our
results suggest that these measures could reduce teen fertility. Moreover, they suggest that these
reductions could be achieved with little or no change in completed fertility. If the costs of teen
fertility are as high as analysts claim, then these policies will generate important social benefits
beyond those typically considered along dimensions such as crime and the labor market.
23
References
Angrist, J.D., and A.B. Krueger (1992). “Estimating the Payoff to Schooling Using the Vietnam-Era
Draft Lottery,” National Bureau of Economic Research, Working Paper No. 4067.
Ashcraft, A. and K. Lang (2006). “The Consequences of Teenage Childbearing,” National Bureau of
Economic Research, Working Paper No. 12485.
Becker, G. (1960). “An Economic Analysis of Fertility,” in Demographic and Economic Change in
Developed Countries, Conference of the Universities-National Bureau Committee for Economic Research,” a
Report of the National Bureau of Economic Research, Princeton, NJ: Princeton University Press.
Becker, G. (1965). “A Theory of the Allocation of Time,” The Economic Journal, 75(299), 493-517.
Becker, G. (1981). “Altruism in the Family and Selfishness in the Market Place,” Economica, 48(189),
1–15.
Becker, G., and H. G. Lewis (1974). "Interaction between quantity and quality of children."
Economics of the family: Marriage, children, and human capital. UMI, 81-90.
Behrman, J. R., M.R. Rosenzweig, & P. Taubman (1994). “Endowments and The Allocation of
Schooling in the Family and in the Marriage Market: The Twins Experiment,” Journal of
Political Economy, 1131-1174.
Berthelon, M. and D. Kruger (2011). “Risky Behavior among Youth: Incapitation Effects of School
on Adolescent Motherhood and Crime in Chile,” Journal of Public Economics 95 (1-2). 41-53.
Black, S. E., P. J. Devereux, and K. G. Salvanes (2008). “Staying In the Classroom and Out of the
Maternity Ward? The Effect of Compulsory Schooling Laws on Teenage Births,” The
Economic Journal, 118(530), 1025–1054.
Breierova, L. and E. Duflo (2004). “The Impact of Education on Fertility and Child Mortality: Do
Fathers Really Matter Less than Mothers?” National Bureau of Economic Research,
Working Paper No. 10513.
Calonico, S., M. D. Cattaneo, and R. Titiunik. (2014). “Robust Nonparametric Confidence Intervals
for Regression-Discontinuity Designs,” Econometrica, forthcoming.
Chevalier, A., and T.K. Viitanen (2003). “The Long-Run Labour Market Consequences of Teenage
Motherhood in Britain,” Journal of Population Economics, 16(2), 323–343.
Clark, D., and H. Royer (2013). “The Effect of Education on Adult Mortality and Health: Evidence
from Britain,” American Economic Review, 103(6), 2087-2120.
Currie, J., and E. Moretti (2003). “Mother’s Education and the Intergenerational Transmission of
Human Capital: Evidence From College Openings,” Quarterly Journal of Economics, 118(4),
1495–1532.
Del Bono, E., & F. Galindo-Rueda (2007). “The Long-Term Impacts of Compulsory Schooling:
Evidence from a Natural Experiment in School Leaving Dates,” Centre for the Economics
of Education, London School of Economics and Political Science Working Paper.
Devereux, P.J., and R.A. Hart (2010). “Forced to be Rich? Returns to Compulsory Schooling in
Britain,” Economic Journal, 120(549), 1345–1364.
Duflo, E., P. Dupas, and M. Kremer (2011). “Education, HIV and Early Fertility: Experimental
Evidence
from
Kenya,”
working
paper
available
online
at
http://www.stanford.edu/~pdupas/DDK_EducFertHIV.pdf.
Geronimus, A. and S. Korenman (1992). “The Socioeconomic Consequences of Teen Childbearing
Reconsidered,” Quarterly Journal of Economics, 107(4), 1187-1214.
Geruso, M (2010). “Education and Childbearing: What are the Links?” Research in Public Policy CMPO Bulletin, University of Bristol.
24
Goldin, C., and L. Katz (2003). “Mass Secondary Schooling and the State: The Role of State
Compulsion in the High School Movement,” National Bureau of Economic Research,
Working Paper No. 10075.
Goodman, A., G. Kaplan, and I. Walker (2004). “Understanding the Effects of Early Motherhood in
Britain: The Effects on Mothers,” Institute for the Study of Labor (IZA) Research Paper
Series, Discussion Paper No. 1131.
Grenet, J. (2013). “Is Extending Compulsory Schooling Alone Enough to Raise Earnings? Evidence
from French and British Compulsory Schooling Laws,” The Scandinavian Journal of Economics,
115: 176–210.
Harmon, C., and I. Walker (1995). “Estimates of the Economic Return to Schooling for the United
Kingdom,” American Economic Review, 85(5), pp. 1278–1286.
Imbens, G.W., and J.D. Angrist (1994). “Identification and Estimation of Local Average Treatment
Effects,” Econometrica, 62(2), pp. 467–475.
Imbens, G. W., and K. Kalyanaraman. (2012). “Optimal Bandwidth Choice for the Regression
Discontinuity Estimator,” Review of Economic Studies 79(3): 933-959.
Imbens, G. W., and T. Lemieux, T. (2008). “Regression Discontinuity Designs: A Guide to
Practice,” Journal of econometrics, 142(2): 615-635.
Kalwij, A. (2010). “The impact of family policy expenditure on fertility in western Europe,”
Demography, 47(2), pp. 503-519.
Lee, D.S., and T. Lemieux (2010). “Regression Discontinuity Designs in Economics,” Journal of
Economic Literature, 48(2), pp. 281–355.
Lefgren, L., and F.L. McIntyre (2006). “The Relationship between Women’s Education and
Marriage Outcomes,” Journal of Labor Economics, 24(4), pp. 787–830.
Lleras-Muney, A. (2005). “The Relationship between Education and Adult Mortality in the United
States,” Review of Economic Studies, 72(1), pp. 189–221.
Ludwig, J., and D.L. Miller (2007). “Does Head Start Improve Children's Life Chances? Evidence
from a Regression Discontinuity Design.” The Quarterly Journal of Economics, 122(1), pp. 159208.
Machin, S., O. Marie, and S. Vuji (2011). “The Crime Reducing Effect of Education,” The Economic
Journal, 121(552), pp. 463-484.
Mazumder, B. (2008). “Does Education Improve Health? A Reexamination of the Evidence from
Compulsory Schooling Laws.” Federal Reserve Bank of Chicago Economic Perspectives, (32) 2:2-16.
McCrary, J., and H. Royer (2011). “The Effect of Female Education on Fertility and Infant Health:
Evidence from School Entry Policies Using Exact Date of Birth,” American Economic Review,
101(1), 158–195.
Mincer, J. (1963), “Market Prices, Opportunity Costs, and Income Effects,” in C. Christ, ed.,
Measurement in Economics: Studies in Mathematical Economics and Econometrics in Memory of Yehuda
Grunfeld, Stanford: Stanford University Press.
Monstad, K., C. Propper, & K. G. Salvanes (2008). “Education and Fertility: Evidence from a
Natural Experiment,” The Scandinavian Journal of Economics, 110(4), pp. 827-852.
Oreopoulos, P. (2006). “Estimating Average and Local Average Treatment Effects of Education
When Compulsory Schooling Laws Really Matter,” The American Economic Review, 96(1), pp.
152–175.
Oreopoulos, P. (2007). “Do Dropouts Drop Out Too Soon? Wealth, Health and Happiness from
Compulsory Schooling,” Journal of Public Economics, 91(11): 2213-2229.
Osili, U.O., and B.T. Long (2008). “Does Female Schooling Reduce Fertility? Evidence from
Nigeria,” Journal of Development Economics, 87(1), pp. 57–75.
25
Painter, G., and D.I. Levine (2004). “Daddies, Devotion, and Dollars: How Do They Matter for
Youth?,” American Journal of Economics and Sociology, 63(4), 813–850.
Preston, S.H., P. Heuveline, and M. Guillot (2001). Demography: Measuring and Modeling Population
Processes, Malden, MA: Blackwell Publishing.
Royer, H. (2004). “What All Women (and Some Men) Want to Know: Does Maternal Age Affect
Infant Health?,” Center for Labor Economics and the University of California, Berkeley,
Working Paper No. 68.
Willis, R. J. (1973). “A New Approach to the Economic Theory of Fertility Behavior,” Journal of
Political Economy, 81(2), pp. S14–S64.
Woodin, T., G. McCulloch, and S. Cowan (2013). Secondary Education and the Raising of the SchoolLeaving Age: Coming of Age? New York, NY: Palgrave MacMillan.
26
APPENDIX
1972 ROSLA
We define the treatment of interest as experiencing an extra year of full-time education at age 15. We
have characterized the 1972 ROSLA as increasing the minimum leaving age from 15 to 16. It seems
natural to suppose that the ROSLA did, therefore, push significant fractions of students into this
treatment.
While this intuition is approximately correct, it is worth noting that the school leaving rules were
slightly more complicated. In particular, there were two possible points at which students could
leave school at the minimum leaving age: Easter (i.e., the end of the spring term) and June/July (i.e.,
the end of the summer term). Students born between 1 September and 31 January were eligible to
leave at Easter; students born 1 Febuary-31 August were required to remain in school until the
summer. For a student born in September 1957, the 1972 ROSLA meant that the earliest leaving
point was Easter 1974 rather than Easter 1973. For a student born in February 1958, the 1972
ROSLA meant that the earliest leaving point was summer 1974 rather than summer 1973.
Because students born between February and August could leave in June/July, some would actually
leave before the minimum school leaving age (both before and after the 1972 ROSLA). This can
account for some of the apparent non-compliance seen in Figure 1. Because the ROSLA required
students to remain until age 16, this meant the earliest academic year they could leave was the
academic year in which students took O Level and CSE exams. Since these exams are taken in
May/June, this meant that some students, such as those born in early July, would technically have
left at the minimum age under the old system (i.e., age 15) but below the minimum age under the
new system (i.e., also at age 15, albeit almost one year later). This can account for some of the
increased non-compliance that appears to follow the 1972 ROSLA.
27
Appendix: Fertility Trends in Europe
It is important to understand the environment in which this change occurred. Fertility declined
dramatically between the mid-1960s and the mid-1970s in the UK. Appendix Figure 1 displays the
trend, along with trends in the US, France and Germany for comparison. In the UK, by 1975, total
fertility had fallen to modern, below-replacement levels.
The post-1972 ROSLA period in the UK, marked by the vertical line, is a period of low and stable
fertility (1972-2002). Modern fertility rates in the UK are quite comparable to those of the US and
France. Given the similarity of fertility today and at the time of the enactment of the 1972 ROSLA,
the 1972 ROSLA is almost ideal for studying the education-fertility relationship: early enough for the
affected women to have now completed their fertility, but late enough for them to be informative
about childbearing in a modern setting. Moreover, the cohorts just affected by the reform have
shared the same access to modern contraception and abortion services throughout their fecund lives
as women today.27
Aside from total fertility, we examine teenage births, often an explicit target of fertility policies. As
seen in Appendix Figure 1, teen fertility also declined through the 1960s and 1970s, although in the
UK it fell slightly slower and slightly later than total fertility. The timing of the fertility decline is
consistent with a possible effect of the 1972 ROSLA. However, all countries depicted in the graph
exhibit a similarly-timed decline, raising the possibility that wider trends could be confounded with a
ROSLA program effect. This highlights the importance of using a research design that can credibly
control for these cohort and time trends. Teen fertility in the UK stabilized at a level higher than
those seen elsewhere in Western Europe but below that seen in the US. As of 2009, about one in
eight girls in the UK bore a child before age 20.
27The 1967 Abortion Act legalized abortion in the UK. This Act covered abortions for teenage girls and did not require
parental consent or notification. Oral contraceptives became available in 1961 through the National Health Service
(http://news.bbc.co.uk/onthisday/hi/dates/stories/december/4/newsid_3228000/3228207.stm).
28
Fraction Leaving School by Different Ages
0
.2
.4
.6
.8
Figure 1: Effect of ROSLA on Age Left Full Time Schooling for Women
1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963
Date of Birth
left school by age 15
left school by age 17
left school by age 16
Notes: The figure depicts the impact of the 1972 ROSLA on educational attainment. Each point represents a mean
for a cohort of women sharing the same month of birth. The plotted lines are local linear regression fits to those
points over the chosen bandwidth. The vertical line indicates the ROSLA threshold, with the first post-reform cohort
born September 1957. Data come from the UK Labor Force Survey, pooled 1975-2002 and restricted to women who
were born in the UK and resident in England or Wales at the time of the interview. Respondents were asked the age
at which they first left continuous, full-time education.
16
Age Left Full-Time Education
16.5
17
17.5
18
Figure 2: Effect of ROSLA on Age at Which Left Full Time Schooling
1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963
Date of Birth
Notes: The impact of the 1972 ROSLA on educational attainment. Each point represents the mean average age at
school exit for a cohort of women sharing the same month of birth. The plotted lines are local linear regression fits
to those points over the chosen bandwidth. The vertical line indicates the ROSLA threshold, with the first postreform cohort born September 1957. Data come from the UK Labor Force Survey, pooled 1975-2002 and restricted
to women who were born in the UK and resident in England or Wales at the time of the interview. Respondents were
asked the age at which they first left continuous, full-time education.
1952
1954
1956 1958 1960
Date of Birth
1962
Fertility Rate at Age 18
1952
1954
1956 1958 1960
Date of Birth
1962
Births Per Woman of Age Group
.02
.03
.04
Fertility Rate at Age 16
Births Per Woman of Age Group
.05 .06 .07 .08 .09
Births Per Woman of Age Group
.04
.05
.06
.07
Births Per Woman of Age Group
.01
.015
.02
Figure 3: Effects of ROSLA on Fertility
Fertility Rate at Age 17
1952
1954
1956 1958 1960
Date of Birth
1962
Fertility Rate at Age 19
1952
1954
1956 1958 1960
Date of Birth
1962
Notes: The impact of the 1972 ROSLA on fertility. Each point represents the fertility rate (number of births/size of
population) for a cohort of women sharing the same month of birth. The plotted lines are local linear regression fits
to those points over the chosen bandwidth. The vertical line indicates the ROSLA threshold, with the first postreform cohort born September 1957.
1954
1956
1958
1960
Date of Birth
1962
Abortion Rate at Age 17
1954
1956
1958
1960
Date of Birth
1962
Abortions Per Woman of Age Group
.005
.01
.015
.02
Abortion Rate at Age 15
Abortions Per Woman of Age Group
.016
.018
.02
.022
Abortions Per Woman of Age Group
.012 .014 .016 .018 .02
Abortions Per Woman of Age Group
.002
.004
.006
.008
Figure 4: Effects of ROSLA on Abortion
Abortion Rate at Age 16
1954
1956
1958
1960
Date of Birth
1962
Abortion Rate at Age 18
1954
1956
1958
1960
Date of Birth
1962
Notes: The impact of the 1972 ROSLA on abortion rates. Each point represents the abortion rate (number of
abortions/size of population) for a cohort of women sharing the same month of birth. The plotted lines are local
linear regression fits to those points over the chosen bandwidth. The vertical line indicates the ROSLA threshold,
with the first post-reform cohort born September 1957.
Figure 5: Effect of ROSLA on Completed Fertility
1.65
Births Per Woman of Age Group
1.7
1.75
1.8
1.85
1.9
Completed Fertility at Age 45
1954
1956
1958
Date of Birth
1960
1962
Notes: The impact of the 1972 ROSLA on completed fertility (i.e, the sum of age specific fertility rates from 16 to
45). Each point represents the average number of children born to a cohort of women sharing the same month of
birth. The plotted lines are local linear regression fits to those points over the chosen bandwidth. The vertical line
indicates the ROSLA threshold, with the first post-reform cohort born September 1957.
Figure 6: Evolution of Cumulative ROSLA Fertility Effects
-.15
-.02
Regression Discontinuity Estimate
-.1
-.05
0
Regression Discontinuity Estimate
-.01
0
.01
.05
ln(Cumulative Fertility) Effects
.02
Cumulative Fertility Effects
15
20
25
30
Age
35
40
45
15
20
25
30
Age
35
40
45
Notes: The impact of the 1972 ROSLA on cumulative fertility. Each point in the plots represents a regression
discontinuity estimate of the reduced form effect of ROSLA on cumulative fertility from 16 up to and including the
indicated age. The regression discontinuity regressions include the month of birth running variable, a dummy for the
post-ROSLA cohorts, and their interaction along with month of birth dummies. Pointwise 95% confidence intervals
are plotted as dotted lines. The left panel plots level effects, which are measured in counts of children per woman.
The right panel plots effects on log of cumulative fertility. Cumulative fertility at each age x is the sum of age-specific
fertility rates from 16 to x, inclusive. Since the first point in each panel is degenerate, representing the cumulative
effect from 16 to 16, it corresponds exactly with the regression discontinuity estimate of the reform on the age 16
fertility rate. Robust standard errors are estimated.
-.4
Ln(Age-Specific Fertility Rate) RD Effects
-.2
0
.2
.4
Figure 7: Effects Over Teen Years
15.5
16
16.5
17
Age
17.5
18
18.5
19
Notes: The impact of the 1972 ROSLA on fertility at certain ages. Each point in the plots represents a regression
discontinuity estimate of the reduced form effect of ROSLA on ln(fertility rate) for that age. The regression
discontinuity regressions include the month of birth running variable, a dummy for the post-ROSLA cohorts, and
their interaction along with month of birth dummies. The green lines mark the start and end of the extra schooling
due to the ROSLA (students can experience it between ages of 14.9 and 16.6 (translating into a birth effect by adding
on 40 weeks of ages 15.6 to 17.3) . Pointwise 95% confidence intervals are plotted as dotted lines. Robust standard
errors are estimated.
Figure 8: Effect of ROSLA on Educational Attainment of Woman's Mate
.5
Fraction with 1+ O-Levels or 1 A-Level
.15
.3
Fraction
.2
.25
Fraction
.35 .4 .45
.3
Fraction with No Qualifications
1952
1954
1956 1958 1960
Female Date of Birth
1962
1952
1954
1956 1958 1960
Female Date of Birth
1962
.25
.26
Fraction
.27 .28
.29
Fraction with 2+ A-Levels or Higher
1952
1954
1956 1958 1960
Female Date of Birth
1962
Notes: The impact of the 1972 ROSLA on the educational attainment of a woman's mate. Each point represents the
fraction of women of the same month of birth cohort who have a mate of a certain education level. The plotted
lines are local linear regression fits to those points over the chosen bandwidth. The vertical line indicates the ROSLA
threshold, with the first post-reform cohort born September 1957.
Figure 9: Effect of ROSLA on Educational Attainment of Man's Mate
Fraction with 1+ O-Levels or 1 A-Level
.1
.4
Fraction
.15
.2
Fraction
.45 .5 .55
.25
.6
Fraction with No Qualifications
1952
1954
1956 1958 1960
Male Date of Birth
1962
1952
1954
1956 1958 1960
Male Date of Birth
1962
.26
Fraction
.27
.28
.29
Fraction with 2+ A-Levels or Higher
1952
1954
1956 1958 1960
Male Date of Birth
1962
Notes: The impact of the 1972 ROSLA on the educational attainment of a man's mate. Each point represents the
fraction of men of the same month of birth cohort who have a mate of a certain education level.The plotted lines
are local linear regression fits to those points over the chosen bandwidth. The vertical line indicates the ROSLA
threshold, with the first post-reform cohort born September 1957.
Figure 10: Effect of ROSLA on Age Gap of Spouses
Wife's Age-Husband's Age
Average Gap in Months
-25
-30
26
Average Gap in Months
27
28
-20
29
Husband's Age-Wife's Age
1952
1954
1956 1958 1960
Female Date of Birth
1962
1952
1954
1956 1958 1960
Male Date of Birth
1962
Notes: The impact of the 1972 ROSLA on age gap of spouses; women's mates are represented by the left panel and
men's mates are represented by the right panel. Each point represents the average age gap adjusted for monthly
seasonality between husband and wife (left panel) and the average age gap between wife and husband (right panel) for
women of the same month of birth cohort (left panel) or men of the same month of birth cohort (right panel).
Seasonality adjustments are formed by regressing the average age gap on a set of month dummies and taking the
residuals and adding the average age gap back. The plotted lines are local linear regression fits to those points over
the chosen bandwidth. The vertical line indicates the ROSLA threshold, with the first post-reform cohort born
September 1957.
Regression Discontinuity Estimate
-.006 -.002 .002 .006
Regression Discontinuity Estimate
-.006 -.002 .002 .006
Figure 11: Effect of ROSLA on Age Gap Between Spouses
Females: Husband's Age-Wife's Age
-20
-15
-10
-5
0
Age Gap in Years
5
10
15
20
10
15
20
Males: Wife's Age-Husband's Age
-20
-15
-10
-5
0
Age Gap in Years
5
Notes: The impact of the 1972 ROSLA on the age gap between spouses; women's mates are represented by the left
panel and men's mates are represented by the right panel. Each point represents a regression discontinuity estimate of
the reduced-form effect of ROSLA on the probability of marrying a mate with a certain age gap (left panel: age
difference between wife and husband for women and right panel: age difference between husband and wife for men).
Regression discontinuity estimates control for seasonality through the inclusion of calendar month of birth dummies.
Pointwise 95% confidence intervals are plotted as dotted lines.
Table 1. Effect of ROSLA on Age of School Exit
Dependent Variable:
left school by age left school by age left school by age
15
16
17
Panel A: Linear cohort trend interacted with threshold
Discontinuity
-0.27**
(0.02)
-0.01
(0.01)
0.31**
(0.06)
Observations
96
96
96
Bandwidth
48
48
48
Mean Dependent Variable
0.437
0.693
0.785
Panel B: Linear cohort trend interacted with threshold + month-of-birth dummies
96
48
16.34
Discontinuity
-0.00
(0.00)
0.24**
(0.04)
Observations
96
96
96
Bandwidth
48
48
48
Mean Dependent Variable
0.437
0.693
0.785
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
96
48
16.34
Discontinuity
0.21**
(0.04)
-0.24**
(0.01)
-0.24**
(0.01)
-0.02*
(0.01)
age left full-time
ed
-0.01
(0.01)
-0.02
(0.01)
-0.01
(0.01)
Observations
72
72
72
72
Bandwidth
48
48
48
48
Mean Dependent Variable
0.437
0.693
0.785
16.34
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Column
headers list the dependent variable. Panel A includes linear controls for cohort trends, interacted with the
threshold to allow different pre- and post-reform trends. Panel B adds dummies for month of birth to
capture seasonality. Panel C drops summer borns to deal with the fact that they could drop out of school
at the end of May of the year they reached the compulsory schooling age. Mean of dependent variable is
measured for the cohort born in August 1957, the last pre-reform cohort. Data come from UK Labour
Force Surveys, 1975-2002. The sample is restricted to women born in the UK and resident in England or
Wales at the time of the interview. Each observation is a cohort-cell mean. Regressions are weighted by
cohort sizes. Robust standard errors in parentheses. Bandwidth is measured in months. *: p-value < 0.05,
**: p-value < 0.01, ***: p-value < 0.001
Table 2. Effect of ROSLA on Teenage Fertility
Dependent Variable: Fertility Rate at Age
15
16
17
18
Panel A: Linear cohort trend interacted with threshold
Discontinuity
-0.0031**
(0.0008)
-0.0013
(0.0012)
-0.0002
(0.0014)
Observations
38
38
38
Bandwidth
19
19
19
Mean Dependent Variable
0.0036
0.0180
0.0377
Panel B: Linear cohort trend interacted with threshold + month-of-birth dummies
38
19
0.0496
38
19
0.0590
-0.0034**
(0.0010)
0.0022
(0.0026)
-0.0004
(0.0014)
Observations
38
38
38
Bandwidth
19
19
19
Mean Dependent Variable
0.0036
0.0180
0.0377
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
38
19
0.0496
38
19
0.0590
Discontinuity
-0.0002
(0.0002)
-0.0005
(0.0005)
-0.0018*
(0.0007)
19
-0.0001
(0.0008)
Discontinuity
-0.0002
(0.0003)
-0.0008
(0.0009)
-0.0024**
(0.0007)
-0.0001
(0.0013)
-0.0001
(0.0018)
Observations
Bandwidth
Mean Dependent Variable
29
19
0.0036
29
19
0.0180
29
19
0.0377
29
19
0.0496
29
19
0.0590
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Column headers list the
dependent variable. Panel A includes linear controls for cohort trends, interacted with the threshold to allow different preand post-reform trends. Panel B adds dummies for month of birth to capture seasonality. Panel C drops summer borns to
deal with the fact that they could drop out of school at the end of May of the year they reached the compulsory schooling
age. Mean of dependent variable is measured for the cohort born in August 1957, the last pre-reform cohort. Data come
from live birth records supplied by the UK Office of National Statistics, containing all births in England and Wales to
mothers born in England, Wales, and Scotland, 1970-2008. These are supplemented with data on cohort sizes generated by
registrar reports of cohort sizes at birth and mortality rates from the Human Mortality Database. Each observation is a
cohort-cell mean. Regressions are weighted by cohort sizes. Robust standard errors in parentheses. Bandwidth is measured
in months. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Table 3. Effect of ROSLA on In/Out of Marriage Teenage Fertility
Dependent Var: In-Marriage Fertility Rate at Age
15
16
17
18
19
Panel A: Linear cohort trend interacted with threshold
Dependent Var: Out-Marriage Fertility Rate at Age
15
16
17
18
19
Discontinuity
-0.0002 -0.0003 -0.0001 -0.0003 0.0003
(0.0002) (0.0004) (0.0004) (0.0005) (0.0005)
0.0000 -0.0015**-0.0031** -0.0010 -0.0004
(0.0000) (0.0004) (0.0006) (0.0009) (0.0011)
Observations
38
38
38
38
38
38
38
38
Bandwidth
19
19
19
19
19
19
19
19
Mean Dep Var
0
0.00842 0.0231 0.0333 0.0453
0.00360 0.00956 0.0146
Panel B: Linear cohort trend interacted with threshold + month-of-birth dummies
Discontinuity
0.0000 -0.0015**-0.0031** -0.0010 -0.0004
(0.0000) (0.0004) (0.0006) (0.0009) (0.0011)
Observations
Bandwidth
Mean Dep Var
0.0000 -0.0011 -0.0023** -0.0007 -0.0005
(0.0000) (0.0006) (0.0006) (0.0011) (0.0014)
29
19
0
29
29
19
19
0.00842 0.0231
29
19
0.0333
29
19
0.0453
38
19
0.0137
-0.0002 0.0002 0.0001 -0.0002 0.0003
(0.0002) (0.0003) (0.0006) (0.0005) (0.0005)
Observations
38
38
38
38
38
38
38
38
Bandwidth
19
19
19
19
19
19
19
19
Mean Dep Var
0
0.00842 0.0231 0.0333 0.0453
0.00360 0.00956 0.0146
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
Discontinuity
38
19
0.0163
38
19
0.0163
38
19
0.0137
-0.0002 0.0003 -0.0001 0.0006 0.0004
(0.0003) (0.0005) (0.0004) (0.0005) (0.0007)
29
29
29
19
19
19
0.00360 0.00956 0.0146
29
19
0.0163
29
19
0.0137
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Column headers list the dependent
variable. Panel A includes linear controls for cohort trends, interacted with the threshold to allow different pre- and post-reform trends.
Panel B adds dummies for month of birth to capture seasonality. Panel C drops summer borns to deal with the fact that they could drop
out of school at the end of May of the year they reached the compulsory schooling age. Mean of dependent variable is measured for the
cohort born in August 1957, the last pre-reform cohort. Data come from live birth records supplied by the UK Office of National
Statistics, containing all births in England and Wales to mothers born in England, Wales, and Scotland, 1970-2008. These are
supplemented with data on cohort sizes generated by registrar reports of cohort sizes at birth and mortality rates from the Human
Mortality Database. Each observation is a cohort-cell mean. Regressions are weighted by cohort sizes. Robust standard errors in
parentheses. Bandwidth is measured in months. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Table 4. Effect of ROSLA on Abortion Rates
15
Panel A: Linear cohort trend interacted with threshold
Discontinuity
-0.00055
(0.00036)
Dependent Variable: Abortion Rate at Age
16
17
18
-0.00002
(0.00057)
0.00033
(0.00042)
Observations
38
38
38
Bandwidth
19
19
19
Mean Dependent Variable
0.00603
0.0152
0.0187
Panel B: Linear cohort trend interacted with threshold + month-of-birth dummies
Discontinuity
-0.00036
(0.00038)
0.00045
(0.00046)
0.00008
(0.00049)
Observations
38
38
38
Bandwidth
19
19
19
Mean Dependent Variable
0.00603
0.0152
0.0187
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
19
0.00071
(0.00058)
-0.00055
(0.00058)
38
19
0.0191
38
19
0.0178
0.00057
(0.00047)
-0.00063
(0.00060)
38
19
0.0191
38
19
0.0178
Discontinuity
-0.00043
(0.00041)
0.00010
(0.00072)
0.00033
(0.00056)
0.00058
(0.00064)
-0.00045
(0.00078)
Observations
Bandwidth
Mean Dependent Variable
29
19
0.00603
29
19
0.0152
29
19
0.0187
29
19
0.0191
29
19
0.0178
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Column headers list the
dependent variable. Panel A includes linear controls for cohort trends, interacted with the threshold to allow different preand post-reform trends. Panel B adds dummies for month of birth to capture seasonality. Panel C drops summer borns to
deal with the fact that they could drop out of school at the end of May of the year they reached the compulsory schooling
age. Mean of dependent variable is measured for the cohort born in August 1957, the last pre-reform cohort. Data come
from abortion records supplied by the UK Office of National Statistics, containing all terminations in England and Wales to
15-19 year old mothers born in England, Wales, and Scotland, 1970-2008. These are supplemented with data on cohort sizes
generated by registrar reports of cohort sizes at birth and mortality rates from the Human Mortality Database. Each
observation is a cohort-cell mean. Regressions are weighted by cohort sizes. Robust standard errors in parentheses.
Bandwidth is measured in months. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Table 5. Effect of ROSLA on Cumulative Fertility
Dependent Variable: Cumulative Fertility Rate for Ages
16-25
16-30
16-35
16-40
Panel A: Linear cohort trend interacted with threshold
Discontinuity
-0.0063
(0.0073)
-0.0005
(0.0073)
Observations
38
Bandwidth
19
Mean Dependent Variable
0.777
Panel B: Linear cohort trend interacted with threshold +
Discontinuity
0.0009
(0.0049)
0.0022
(0.0059)
38
38
19
19
1.327
1.639
month-of-birth dummies
-0.0011
(0.0062)
-0.0075
(0.0070)
Observations
38
38
38
Bandwidth
19
19
19
Mean Dependent Variable
0.777
1.327
1.639
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
Discontinuity
Observations
Bandwidth
Mean Dependent Variable
16-45
0.0037
(0.0058)
0.0036
(0.0058)
38
19
1.756
38
19
1.776
-0.0050
(0.0084)
-0.0055
(0.0089)
38
19
1.756
38
19
1.776
-0.0004
(0.0093)
0.0058
(0.0092)
0.0075
(0.0080)
0.0074
(0.0080)
0.0073
(0.0080)
29
19
0.777
29
19
1.327
29
19
1.639
29
19
1.756
29
19
1.776
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Cumulative fertility at each
age x is the sum of age-specific fertility rates 16 to x, inclusive. Column headers list the dependent variable. Panel A includes
linear controls for cohort trends, interacted with the threshold to allow different pre- and post-reform trends. Panel B adds
dummies for month of birth to capture seasonality. Panel C drops summer borns to deal with the fact that they could drop
out of school at the end of May of the year they reached the compulsory schooling age. Mean of dependent variable is
measured for the cohort born in August 1957, the last pre-reform cohort. Data come from live birth records supplied by the
UK Office of National Statistics, containing all births in England and Wales to mothers born in England, Wales, and Scotland,
1970-2008. These are supplemented with data on cohort sizes generated by registrar reports of cohort sizes at birth and
mortality rates from the Human Mortality Database. Each observation is a cohort-cell mean. Regressions are weighted by
cohort sizes. Robust standard errors in parentheses. Bandwidth is measured in months. * :p-value < 0.05, ** :p-value < 0.01,
***: p-value < 0.001
Table 6. Partnerships and Preferences at Age 16
Left School at 16
Preferred 15 Preferred 16
Left at 17+
Panel A
Lived with guardians on 16th birthday
Left home at age 16
Left home at age 17
0.99
0.10
0.23
0.99
0.08
0.19
1.00
0.02
0.09
First cohabitating partnership at 16
First cohabitating partnership at 17
Obs
0.06
0.11
1039
0.02
0.07
1323
0.00
0.02
1842
18
20
17.07
0.84
18.33
0.36
0.66
0.79
61
18
19.5
13.99
0.79
20.59
0.32
0.54
0.54
33
46
46
12.35
0.60
26.96
0.00
0.00
0.00
5
19
20
15.89
0.96
18.30
0.05
0.36
0.64
117
19
21
17.23
0.88
19.87
0.06
0.41
0.59
86
19
21
14.78
0.69
22.08
0.00
0.23
0.45
32
Panel B
Respondents who cohabitated at 16
Parnter age: 25th percentile
Parnter age: Median
Duration of partnership
Eventual marriage
Age at first birth
Birth by 16
Birth by 17
Birth by 18
Obs
Respondents who cohabitated at 17
Parnter age: 25th percentile
Parnter age: Median
Duration of partnership
Eventual marriage
Age at first birth
Birth by 16
Birth by 17
Birth by 18
Obs
Notes: Sample includes NCDS female respondents. Columns sub-divide sample by preferred schoolleaving age, as reported by the NCDS cohort member at the time they were 16. Respondents were asked:
"... do you wish that you could have left when you were 15?" Missing cohort member responses are filled
in with their parents' responses to the question: "In the study child’s case do they wish that he/she had
been able to leave school at fifteen?" If both parent and respondent answered "unsure," respondent is
coded with a preference for leaving at age 15.
Table 7. Effect of ROSLA on Partner's Educational Credentials
Woman's Mate
Man's Mate
No
1+ O-levels 2+ A-Levels
Qualifications or 1 A-level or Higher
Panel A: Linear cohort trend interacted with threshold
Discontinuity
-0.0122**
(0.0022)
0.0123**
(0.0016)
0.0036
(0.0025)
Observations
72
72
Bandwidth
36
36
Mean Dep Var
0.237
0.361
Panel B: Linear cohort trend interacted with threshold +
Discontinuity
-0.0123**
(0.0019)
0.0124**
(0.0018)
No
1+ O-levels 2+ A-Levels
Qualifications or 1 A-level or Higher
-0.0137**
(0.0022)
72
72
36
36
0.266
0.187
month-of-birth dummies
0.0028
(0.0021)
-0.0143**
(0.0019)
Observations
72
72
72
72
Bandwidth
36
36
36
36
Mean Dep Var
0.237
0.361
0.266
0.187
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
0.0125**
(0.0021)
0.0028
(0.0026)
72
36
0.516
72
36
0.261
0.0140**
(0.0017)
0.0017
(0.0019)
72
36
0.516
72
36
0.261
Discontinuity
-0.0101**
(0.0026)
0.0131**
(0.0020)
0.0016
(0.0027)
-0.0104**
(0.0026)
0.0155**
(0.0020)
-0.0034
(0.0023)
Observations
Bandwidth
Mean Dep Var
54
36
0.237
54
36
0.361
54
36
0.266
54
36
0.187
54
36
0.516
54
36
0.261
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Column headers list the
dependent variable. Panel A includes linear controls for cohort trends, interacted with the threshold to allow different preand post-reform trends. Panel B adds dummies for month of birth to capture seasonality. Panel C drops summer borns to
deal with the fact that they could drop out of school at the end of May of the year they reached the compulsory schooling
age. Mean of dependent variable (Mean Dep Var) is measured for the cohort born in August 1957, the last pre-reform
cohort. Data come from the 2001 Census. Each observation is a cohort-cell mean. Regressions are weighted by cohort sizes.
Robust standard errors in parentheses. Bandwidth is measured in months. * :p-value < 0.05, ** :p-value < 0.01, ***: pvalue < 0.001
Appendix Figure 1: Global Trends in Fertility During Study Period
1
2
TFR
3
4
Total Fertility
1940
1950
1970
1980
year
1990
2000
US
Germany
Teen Fertility
2010
France
0
births per 1000 women aged 15−19
20
40
60
80
100
UK
1960
1960
1970
1980
1990
2000
2010
year
UK
US
Germany
France
Notes: Total fertility rates in the UK, US, France, and Germany. The vertical line indicates the date
of the 1972 ROSLA. The raising of the school leaving age in 1972 came at the tail end of a
demographic transition among industrialized nations. Women just affected by the reform
experienced their childbearing years (roughly 1972 through 2002) in a period of low, stable fertility.
UK data come from the UK Office of National Statistics. US, French, and German data come from
the Human Fertility Database.
1952
1954
1956 1958 1960
Date of Birth
1962
Out of Marriage Fertility: Age 16
1952
1954
1956 1958 1960
Date of Birth
1962
Births Per Woman of Age Group
.01 .015 .02 .025 .03
In Marriage Fertility: Age 16
Births Per Woman of Age Group
.01 .015 .02 .025 .03
Births Per Woman of Age Group
.002 .004 .006 .008 .01
Births Per Woman of Age Group
.002 .004 .006 .008 .01
Appendix Figure 2: Effect of ROSLA on In and Out of Marriage Fertility
In Marriage Fertility: Age 17
1952
1954
1956 1958 1960
Date of Birth
1962
Out of Marriage Fertility: Age 17
1952
1954
1956 1958 1960
Date of Birth
1962
Notes: The impact of the 1972 ROSLA on in marriage and out of marriage fertility. Each point represents the
fertility rate (number of births/size of age group) for a cohort of women sharing the same month of birth. The
plotted lines are local linear regression fits to those points over the chosen bandwidth. The vertical line indicates the
ROSLA threshold, with the first post-reform cohort born September 1957.
NCDS: Fraction wish left before 16
.15
.2
.25
.3
.35
.4
.45
Appendix Figure 3: Self-Reports of Complier Status
.15
.2
.25
.3
.35
.4
.45
LFS: Fraction left school before 16
men
women
NCDS location is place of birth.
Notes: The vertical axis
plotslocation
the fraction is
wishing
to leave
age 16, self-reported by NCDS
LFS
place
of school
adultbefore
residence
respondents (born in the post-ROSLA period). The horizontal axis plots the fraction actually exiting from school
NCDS cohort born March 1958
before age 16, self-reported by LFS respondents who were born in the pre-period 1954-1957. Each point
LFS born 1954-1957
corresponds to a region of the UK interacted with gender.
Appendix Figure 4: Effect of ROSLA on Probability of Marrying Someone in
the Same Academic Cohort
Fraction Marrying a Classmate
.11
.12
.13
.1
.1
Fraction Marrying a Classmate
.11
.12
.13
.14
Males
.14
Females
1952
1954
1956 1958 1960
Female Date of Birth
1962
1952
1954
1956 1958 1960
Male Date of Birth
1962
Notes: The impact of the 1972 ROSLA on probability of marrying someone in the same academic cohort; women's
mates are represented by the left panel and men's mates are represented by the right panel. Each point represents the
fraction marrying a classmate for women of the same month of birth cohort (left panel) or men of the same month
of birth cohort (right panel). Seasonality adjustments are formed by regressing the cohort fraction on a set of month
of birth dummies and taking the residuals and adding the mean fraction marrying a classmate. The plotted lines are
local linear regression fits to those points over the chosen bandwidth. The vertical line indicates the ROSLA
threshold, with the first post-reform cohort born September 1957.
Appendix Figure 5: Effect of ROSLA on Probability of Marrying Someone in
the Post-ROSLA Cohort
Fraction Marrying Someone Born September 1957 or Later
.2
.4
.6
.8
0
0
Fraction Marrying Someone Born September 1957 or Later
.2
.4
.6
.8
1
Males
1
Females
1952
1954
1956 1958 1960
Female Date of Birth
1962
1952
1954
1956 1958 1960
Male Date of Birth
1962
Notes: The impact of the 1972 ROSLA on probability of marrying someone born September 1957 or later; women's
mates are represented by the left panel and men's mates are represented by the right panel. Each point represents the
fraction marrying someone born September 1957 of later for women of the same month of birth cohort (left panel)
or men of the same month of birth cohort (right panel). Seasonality adjustments are formed by regressing the cohort
fraction on a set of month of birth dummies and taking the residuals and adding the mean fraction marrying
someone born September 1957 or later. The plotted lines are local linear regression fits to those points over the
chosen bandwidth. The vertical line indicates the ROSLA threshold, with the first post-reform cohort born
September 1957.
Appendix Table 1: Data Sources
Data Source
Sample
UK Labour Force Sample of
Survey (LFS)
households in
England and Wales
Primary Variables
age left full-time
education
Notes
individual-level
Live Births Records 100% sample of
live births
Extract
births in England and
Wales 1970-2008
counts by mother's date of birth
(measured in months) and child's date
of birth (measured in months); not an
individual-level panel
Abortion Records
Extract
counts by mother's date of birth
(measured in months) and date of
procedure (measured in months)
100% sample of teen abortions
abortions (15-19
years old) in England
and Wales
UK Census Extracts 100% sample for the marital status, partner's
2001 Census
age, partner's education
National Child
100% sample of
Development Study children born one
(1958 Cohort)
week in March 1978
age at which would have one of the first post-ROSLA cohorts
dropped out of school if
ROSLA had not
occurred, activities in
school, partnership
histories, activities out of
school
Appendix Table 2. Sensitivity of Regression Discontinuity Estimates to Varying Bandwidths
Outcome
Left School by Age 15
Chosen
Bandwidth in
Chosen
Months
Bandwidth - 6
48
-0.24**
(0.01)
Chosen
Bandwidth
-0.24**
(0.01)
Chosen
Bandwidth + 6
-0.24**
(0.01)
Left School by Age 16
48
-0.02*
(0.01)
-0.01
(0.01)
-0.01
(0.01)
Left School by Age 17
48
-0.01
(0.01)
-0.00
(0.00)
-0.00
(0.00)
Age Left Schooling
48
0.26**
(0.04)
0.24**
(0.04)
0.21**
(0.03)
Fertility at Age 15
19
-0.0002
(0.0003)
-0.0002
(0.0003)
-0.0001
(0.0002)
Fertility at Age 16
19
-0.0013
(0.0010)
-0.0009
(0.0006)
-0.0012*
(0.0005)
Fertility at Age 17
19
-0.0011
(0.0008)
-0.0021**
(0.0007)
-0.0030**
(0.0006)
Fertility at Age 18
19
-0.0002
(0.0012)
-0.0008
(0.0011)
-0.0021
(0.0011)
Fertility at Age 19
19
0.0021
0.0004
-0.0015
(0.0011)
(0.0013)
(0.0012)
Notes: Notes: Each column and row is the regression discontinuity estimate from a separate regression. All
regression estimates come from local linear regressions which allow the slope with respect to the running
variable to vary pre and post reform. These regressions also include month-of-birth dummies. Robust
standard errors in parentheses. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Appendix Table 2 Con't. Sensitivity of Regression Discontinuity Estimates to Varying Bandwidths
Chosen
Bandwidth in
Chosen
Chosen
Chosen
Outcome
Months
Bandwidth - 6 Bandwidth Bandwidth + 6
In-Marriage Fertility at Age 15
19
0.0000
0.0000
0.0000
(0.0000)
(0.0000)
(0.0000)
In-Marriage Fertility at Age 16
19
-0.0014*
(0.0006)
-0.0011**
(0.0003)
-0.0013**
(0.0003)
In-Marriage Fertility at Age 17
19
-0.0014*
(0.0005)
-0.0022**
(0.0005)
-0.0028**
(0.0005)
In-Marriage Fertility at Age 18
19
0.0014
(0.0011)
-0.0005
(0.0010)
-0.0018
(0.0009)
In-Marriage Fertility at Age 19
19
0.0021
(0.0010)
0.0001
(0.0011)
-0.0013
(0.0010)
Out-of-Marriage Fertility at Age 15
19
-0.0002
(0.0003)
-0.0002
(0.0002)
-0.0001
(0.0002)
Out-of-Marriage Fertility at Age 16
19
0.0001
(0.0005)
0.0002
(0.0003)
0.0002
(0.0003)
Out-of-Marriage Fertility at Age 17
19
0.0002
(0.0008)
0.0001
(0.0006)
-0.0002
(0.0004)
Out-of-Marriage Fertility at Age 18
19
-0.0016**
(0.0005)
-0.0002
(0.0005)
-0.0004
(0.0005)
Out-of-Marriage Fertility at Age 19
19
-0.0000
0.0003
-0.0002
(0.0003)
(0.0005)
(0.0004)
Notes: Notes: Each column and row is the regression discontinuity estimate from a separate regression. All
regression estimates come from local linear regressions which allow the slope with respect to the running
variable to vary pre and post reform. These regressions also include month-of-birth dummies. Robust
standard errors in parentheses. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Appendix Table 2 Con't. Sensitivity of Regression Discontinuity Estimates to Varying Bandwidths
Chosen
Bandwidth in
Chosen
Chosen
Chosen
Outcome
Months
Bandwidth - 6 Bandwidth Bandwidth + 6
Abortion at Age 15
19
-0.00031
-0.00036
-0.00047
(0.00041)
(0.00038)
(0.00031)
Abortion at Age 16
19
-0.00087
(0.00056)
0.00045
(0.00046)
-0.00024
(0.00049)
Abortion at Age 17
19
-0.00105
(0.00065)
0.00008
(0.00049)
-0.00001
(0.00041)
Abortion at Age 18
19
0.00018
(0.00066)
0.00057
(0.00047)
0.00020
(0.00044)
Abortion at Age 19
19
-0.00043
(0.00079)
-0.00063
(0.00060)
-0.00094
(0.00057)
Cumulative Fertility Ages 16-25
19
-0.0017
(0.0062)
-0.0004
(0.0055)
-0.0108
(0.0057)
Cumulative Fertility Ages 16-30
19
0.0003
(0.0070)
0.0031
(0.0059)
-0.0094
(0.0065)
Cumulative Fertility Ages 16-35
19
0.0036
(0.0070)
0.0031
(0.0056)
-0.0087
(0.0062)
Cumulative Fertility Ages 16-40
19
0.0025
(0.0064)
0.0041
(0.0058)
-0.0079
(0.0062)
Cumulative Fertility Ages 16-45
19
0.0006
0.0037
-0.0084
(0.0066)
(0.0059)
(0.0064)
Notes: Notes: Each column and row is the regression discontinuity estimate from a separate regression. All
regression estimates come from local linear regressions which allow the slope with respect to the running
variable to vary pre and post reform. These regressions also include month-of-birth dummies. Robust
standard errors in parentheses. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Appendix Table 2 Con't. Sensitivity of Regression Discontinuity Estimates to Varying Bandwidths
Chosen
Bandwidth in
Chosen
Chosen
Chosen
Outcome
Months
Bandwidth - 6 Bandwidth Bandwidth + 6
Woman's Mate: No Qualifications
36
-0.0112**
-0.0123**
-0.0118**
(0.0021)
(0.0019)
(0.0017)
Woman's Mate: 1+ O-levels or 1 A-level
36
0.0117**
(0.0020)
0.0124**
(0.0018)
0.0148**
(0.0017)
Woman's Mate: 2+ A-Levels or Higher
36
0.0029
(0.0022)
0.0028
(0.0021)
0.0007
(0.0018)
Man's Mate: No Qualifications
36
-0.0130**
(0.0019)
-0.0143**
(0.0019)
-0.0119**
(0.0018)
Man's Mate: 1+ O-levels or 1 A-level
36
0.0124**
(0.0018)
0.0140**
(0.0017)
0.0139**
(0.0015)
Man's Mate: 2+ A-Levels or Higher
36
0.0019
0.0017
0.0000
(0.0018)
(0.0019)
(0.0016)
Notes: Notes: Each column and row is the regression discontinuity estimate from a separate regression. All
regression estimates come from local linear regressions which allow the slope with respect to the running
variable to vary pre and post reform. These regressions also include month-of-birth dummies. Robust
standard errors in parentheses. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001
Appendix Table 3. Effect of ROSLA on Conception Rates
Dependent Variable: Conception Rate at Age of Time of Birth
16
17
18
19
Panel A: Linear cohort trend interacted with threshold
Discontinuity
-0.0021
(0.0012)
-0.0020
(0.0016)
Observations
48
48
48
Bandwidth
24
24
24
Mean Dependent Variable
0.0275
0.0548
0.0684
Panel B: Linear cohort trend interacted with threshold + month-of-birth dummies
48
24
0.0772
Discontinuity
-0.0033**
(0.0007)
-0.0014
(0.0012)
-0.0010
(0.0016)
Observations
48
48
48
Bandwidth
24
24
24
Mean Dependent Variable
0.0275
0.0548
0.0684
Panel C: Linear cohort trend interacted with threshold with summer borns dropped
48
24
0.0772
Discontinuity
Observations
Bandwidth
Mean Dependent Variable
-0.0011
(0.0007)
-0.0050**
(0.0008)
-0.0028**
(0.0009)
-0.0023*
(0.0010)
-0.0042**
(0.0009)
-0.0014
(0.0014)
-0.0012
(0.0018)
36
24
0.0275
36
24
0.0548
36
24
0.0684
36
24
0.0772
Notes: Each column and row is the regression discontinuity estimate from a separate regression. Conception rates are
defined as the sum of birth rates and abortion rates (adjusted to take into account that timing of abortion and births are
different, see text for more details). Column headers list the dependent variable. Panel A includes linear controls for
cohort trends, interacted with the threshold to allow different pre- and post-reform trends. Panel B adds dummies for
month of birth to capture seasonality. Panel C drops summer borns to deal with the fact that they could drop out of
school at the end of May of the year they reached the compulsory schooling age. Mean of dependent variable is
measured for the cohort born in August 1957, the last pre-reform cohort. Data come from live birth records and abortion
records supplied by the UK Office of National Statistics, containing all births in England and Wales to mothers born in
England, Wales, and Scotland, 1970-2008. These are supplemented with data on cohort sizes from the Human Mortality
Database. Each observation is a cohort-cell mean. Regressions are weighted by cohort sizes. Robust standard errors in
parentheses. * :p-value < 0.05, ** :p-value < 0.01, ***: p-value < 0.001